Method: Proposes OmniForcing framework that distills offline dual-stream bidirectional diffusion models into streaming autoregressive generators. Key innovations include: 1) Asymmetric Block-Causal Alignment with zero-truncation Global Prefix to handle modality asymmetry, 2) Audio Sink Token mechanism with Identity RoPE constraint to address audio token sparsity, 3) Joint Self-Forcing Distillation to correct cumulative cross-modal errors from exposure bias, and 4) modality-independent rolling KV-cache inference scheme.

Result: Achieves state-of-the-art streaming generation at ~25 FPS on a single GPU while maintaining multi-modal synchronization and visual quality comparable to the bidirectional teacher model. Successfully addresses training instability issues from modality asymmetry and token sparsity.

Conclusion: OmniForcing enables real-time, high-fidelity audio-visual generation by effectively distilling bidirectional diffusion models into efficient streaming autoregressive generators, solving key challenges in multi-modal temporal alignment and training stability.

Abstract: Recent joint audio-visual diffusion models achieve remarkable generation quality but suffer from high latency due to their bidirectional attention dependencies, hindering real-time applications. We propose OmniForcing, the first framework to distill an offline, dual-stream bidirectional diffusion model into a high-fidelity streaming autoregressive generator. However, naively applying causal distillation to such dual-stream architectures triggers severe training instability, due to the extreme temporal asymmetry between modalities and the resulting token sparsity. We address the inherent information density gap by introducing an Asymmetric Block-Causal Alignment with a zero-truncation Global Prefix that prevents multi-modal synchronization drift. The gradient explosion caused by extreme audio token sparsity during the causal shift is further resolved through an Audio Sink Token mechanism equipped with an Identity RoPE constraint. Finally, a Joint Self-Forcing Distillation paradigm enables the model to dynamically self-correct cumulative cross-modal errors from exposure bias during long rollouts. Empowered by a modality-independent rolling KV-cache inference scheme, OmniForcing achieves state-of-the-art streaming generation at $\sim$25 FPS on a single GPU, maintaining multi-modal synchronization and visual quality on par with the bidirectional teacher.\textbf{Project Page:} \href{https://omniforcing.com}{https://omniforcing.com}

Method: SPARROW uses two key components: (1) Target-Specific Tracked Features (TSF) that inject temporally aligned referent cues during training, and (2) a dual-prompt design that decodes box and segmentation tokens to fuse geometric priors with semantic grounding. It operates end-to-end without external detectors via a class-agnostic SAM2-based proposer.

Result: Integrated into three open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks: up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG.

Conclusion: SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding, demonstrating effective unification of spatial accuracy and temporal stability in video MLLMs.

Abstract: Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: https://risys-lab.github.io/SPARROW

Method: Proposes TASTE-S with two key innovations: 1) Integrates a CTC-based ASR module into the encoder for instant dual-modality encoding, eliminating need for external ASR, 2) Redesigns the unit decoder to enable on-the-fly decoding for streaming capabilities. Uses joint training approach.

Result: TASTE-S matches TASTE’s performance while significantly reducing latency. Further investigations show TASTE-S remains robust to transcriptions and enables long-form encoding and decoding capabilities.

Conclusion: TASTE-S successfully addresses the streaming limitation of previous text-speech joint SLM approaches, providing a practical solution for real-time speech-based interactions while maintaining performance and enabling long-form processing.

Abstract: Text-speech joint spoken language modeling (SLM) aims at natural and intelligent speech-based interactions, but developing such a system may suffer from modality mismatch: speech unit sequences are much longer than text tokens. Prior work reduces this gap with text-aligned tokenization and embedding (TASTE), producing speech tokens that align in lengths with their textual counterparts. However, the dependence on an external ASR system and the use of a non-causal decoder limits streaming use. To address this limitation, we propose TASTE-S, a streamable extension of TASTE suitable for real-time usage. TASTE-S integrates a CTC-based ASR module into the encoder for instant dual-modality encoding. We also redesign the unit decoder to enable on-the-fly decoding. With joint training, we show that TASTE-S matches TASTE’s performance while significantly reducing latency. Further investigations reveal that TASTE-S remains robust to transcriptions and enables long-form encoding and decoding.

Method: Introduces a distillation framework that trains lightweight probes on frozen teacher hidden states and uses the probe’s predictions (rather than output logits) as supervision for student training. This bypasses the vocabulary projection bottleneck.

Result: Consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes provide cleaner labels than teacher’s own outputs, effectively denoising the distillation signal.

Conclusion: The method enables practitioners to extract more value from large teacher models without additional training data or architectural complexity by exploiting internal representations, requiring no architectural changes and adding minimal compute.

Abstract: Knowledge distillation from large language models (LLMs) assumes that the teacher’s output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model’s intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe’s predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher’s own outputs, effectively denoising the distillation signal. \method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.

Method: Introduced Dynamic Knowledge Instance (DKI) framework modeling multi-updates as cue-value sequences, evaluated via endpoint probing of earliest and latest states. Used diagnostic analyses of attention patterns, hidden-state similarity, and output logits across diverse LLMs.

Result: Retrieval bias intensifies with more updates: earliest-state accuracy remains high while latest-state accuracy drops substantially. Diagnostic signals become flatter and weakly discriminative on errors. Cognitive-inspired interventions yield only modest gains.

Conclusion: LLMs face persistent challenges in tracking knowledge updates in long contexts, with retrieval bias that cognitive-inspired strategies cannot fully eliminate, revealing fundamental limitations in dynamic knowledge management.

Abstract: LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.

Method: Proposes ActTail with TopK magnitude-based activation sparsity using global allocation based on Heavy-Tailed Self-Regularization theory. Computes heavy-tail exponent from each projection’s empirical spectral density to assign projection-specific sparsity budgets. Provides theoretical analysis linking activation sparsity ratio to heavy-tail exponent.

Result: Experiments on LLaMA and Mistral models show improved perplexity and downstream task performance at high sparsity compared to uniform allocation. At 80% sparsity: 21.8% perplexity reduction on LLaMA-2-7B, 40.1% on LLaMA-2-13B, and 9.4% on Mistral-7B.

Conclusion: ActTail provides principled guidance for sparsity allocation beyond heuristic design, effectively accelerating LLM inference while minimizing performance degradation by accounting for weight heterogeneity through heavy-tail exponent analysis.

Abstract: Activation sparsity is a promising approach for accelerating large language model (LLM) inference by reducing computation and memory movement. However, existing activation sparsity methods typically apply uniform sparsity across projections, ignoring the heterogeneous statistical properties of Transformer weights and thereby amplifying performance degradation. In this paper, we propose ActTail, a TopK magnitude-based activation sparsity method with global activation sparsity allocation grounded in Heavy-Tailed Self-Regularization (HT-SR) theory. Specifically, we capture this heterogeneity via the heavy-tail exponent computed from each projection’s empirical spectral density (ESD), which is used as a quantitative indicator to assign projection-specific sparsity budgets. Importantly, we provide a theoretical analysis that establishes an explicit relationship between the activation sparsity ratio and the heavy-tail exponent under the HT-SR regime, offering principled guidance for sparsity allocation beyond heuristic design. Experiments on LLaMA and Mistral models show that our method improves both perplexity and downstream task performance at high sparsity compared to uniform allocation. At 80% sparsity, perplexity is reduced by 21.8% on LLaMA-2-7B, 40.1% on LLaMA-2-13B, and 9.4% on Mistral-7B.

Method: Self-distillation approach: condition the model on user follow-up messages to obtain hindsight token distributions, then distill these distributions back into the current policy. This captures how the model would revise its behavior after seeing user feedback.

Result: Training on real-world WildChat conversations improves language models across standard alignment and instruction-following benchmarks without regressing other capabilities. The method also enables personalization through continual adaptation to individual users.

Conclusion: Raw user interactions during deployment provide a scalable resource for alignment, personalization, and continual adaptation through self-distillation from hindsight reasoning.

Abstract: Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user’s preferences. Importantly, language models are already able to make use of this information in context. After observing a user’s follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user’s follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model’s behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.

Method: The paper introduces GONE benchmark for evaluating KG-based knowledge unlearning and proposes NEDS (Neighborhood-Expanded Distribution Shaping), a novel unlearning framework that leverages graph connectivity to identify anchor correlated neighbors and enforce precise decision boundaries between forgotten facts and their semantic neighborhoods.

Result: Evaluations on LLaMA-3-8B and Mistral-7B show NEDS achieves superior performance with 1.000 on unlearning efficacy and 0.839 on locality on GONE benchmark, outperforming other knowledge editing and unlearning methods.

Conclusion: GONE benchmark enables better evaluation of knowledge unlearning over structured knowledge, and NEDS provides an effective framework for precise unlearning while preserving semantic neighborhoods, addressing limitations of existing methods.

Abstract: Unlearning knowledge is a pressing and challenging task in Large Language Models (LLMs) because of their unprecedented capability to memorize and digest training data at scale, raising more significant issues regarding safety, privacy, and intellectual property. However, existing works, including parameter editing, fine-tuning, and distillation-based methods, are all focused on flat sentence-level data but overlook the relational, multi-hop, and reasoned knowledge in naturally structured data. In response to this gap, this paper introduces Graph Oblivion and Node Erasure (GONE), a benchmark for evaluating knowledge unlearning over structured knowledge graph (KG) facts in LLMs. This KG-based benchmark enables the disentanglement of three effects of unlearning: direct fact removal, reasoning-based leakage, and catastrophic forgetting. In addition, Neighborhood-Expanded Distribution Shaping (NEDS), a novel unlearning framework, is designed to leverage graph connectivity and identify anchor correlated neighbors, enforcing a precise decision boundary between the forgotten fact and its semantic neighborhood. Evaluations on LLaMA-3-8B and Mistral-7B across multiple knowledge editing and unlearning methods showcase NEDS’s superior performance (1.000 on unlearning efficacy and 0.839 on locality) on GONE and other benchmarks. Code is available at https://anonymous.4open.science/r/GONE-4679/.

Method: Design novel role probes to capture how models internally identify “who is speaking.” Test the insight by injecting spoofed reasoning into user prompts and tool outputs. Evaluate across multiple open- and closed-weight models with near-zero baselines.

Result: Achieved average success rates of 60% on StrongREJECT and 61% on agent exfiltration. Found that degree of internal role confusion strongly predicts attack success before generation begins. Demonstrated that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.

Conclusion: Reveals fundamental gap: security is defined at the interface but authority is assigned in latent space. Introduces unifying mechanistic framework showing prompt injection attacks exploit role confusion where untrusted text imitating a role inherits that role’s authority.

Abstract: Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify “who is speaking.” These reveal why prompt injection works: untrusted text that imitates a role inherits that role’s authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.

Method: Curated 5,059 Reddit posts from mental health subreddits (2010-2025), extracted 23,399 medication mentions, developed aspect-based sentiment classifier by fine-tuning DeBERTa-v3 on SMM4H 2023 Twitter corpus with LLM-based data augmentation, then applied to Reddit data.

Result: 72.1% of medication mentions were neutral, 14.8% negative, 13.1% positive. Conventional antidepressants (SSRIs/SNRIs) showed higher negative than positive sentiment, while ketamine/esketamine had more favorable sentiment profiles.

Conclusion: Normalized medication extraction with aspect-based sentiment analysis can characterize patient-perceived treatment experiences in TRD discourse, providing complementary patient-generated perspectives to clinical evidence.

Abstract: Treatment-resistant depression (TRD) is a severe form of major depressive disorder in which patients do not achieve remission despite multiple adequate treatment trials. Evidence across pharmacologic options for TRD remains limited, and trials often do not fully capture patient-reported tolerability. Large-scale online peer-support narratives therefore offer a complementary lens on how patients describe and evaluate medications in real-world use. In this study, we curated a corpus of 5,059 Reddit posts explicitly referencing TRD from 3,480 subscribers across 28 mental health-related subreddits from 2010 to 2025. Of these, 3,839 posts mentioned at least one medication, yielding 23,399 mentions of 81 generic-name medications after lexicon-based normalization of brand names, misspellings, and colloquialisms. We developed an aspect-based sentiment classifier by fine-tuning DeBERTa-v3 on the SMM4H 2023 therapy-sentiment Twitter corpus with large language model based data augmentation, achieving a micro-F1 score of 0.800 on the shared-task test set. Applying this classifier to Reddit, we quantified sentiment toward individual medications across three categories: positive, neutral, and negative, and tracked patterns by drug, subscriber, subreddit, and year. Overall, 72.1% of medication mentions were neutral, 14.8% negative, and 13.1% positive. Conventional antidepressants, especially SSRIs and SNRIs, showed consistently higher negative than positive proportions, whereas ketamine and esketamine showed comparatively more favorable sentiment profiles. These findings show that normalized medication extraction combined with aspect-based sentiment analysis can help characterize patient-perceived treatment experiences in TRD-related Reddit discourse, complementing clinical evidence with large-scale patient-generated perspectives.

Method: Proposes TASTE-S with two key innovations: 1) Integrates a CTC-based ASR module into the encoder for instant dual-modality encoding, eliminating need for external ASR, 2) Redesigns the unit decoder to enable on-the-fly decoding for streaming capabilities. Uses joint training approach.

Result: TASTE-S matches TASTE’s performance while significantly reducing latency. Further investigations show TASTE-S remains robust to transcriptions and enables long-form encoding and decoding capabilities.

Conclusion: TASTE-S successfully addresses the streaming limitation of previous text-speech joint SLM approaches, providing a practical solution for real-time speech-based interactions while maintaining performance and enabling long-form processing.

Abstract: Text-speech joint spoken language modeling (SLM) aims at natural and intelligent speech-based interactions, but developing such a system may suffer from modality mismatch: speech unit sequences are much longer than text tokens. Prior work reduces this gap with text-aligned tokenization and embedding (TASTE), producing speech tokens that align in lengths with their textual counterparts. However, the dependence on an external ASR system and the use of a non-causal decoder limits streaming use. To address this limitation, we propose TASTE-S, a streamable extension of TASTE suitable for real-time usage. TASTE-S integrates a CTC-based ASR module into the encoder for instant dual-modality encoding. We also redesign the unit decoder to enable on-the-fly decoding. With joint training, we show that TASTE-S matches TASTE’s performance while significantly reducing latency. Further investigations reveal that TASTE-S remains robust to transcriptions and enables long-form encoding and decoding.

Method: Controlled experiment holding harmful final answers constant while varying reasoning paths (Evil, Misleading, Submissive). Trained models (0.6B-14B parameters) under QTA, QT, and T-only paradigms, evaluated in think and no-think modes.

Result: CoT training amplifies harmful generalization more than standard fine-tuning; distinct reasoning types induce distinct behavioral patterns despite identical answers; reasoning-only training alters behavior; effects persist without reasoning generation.

Conclusion: Reasoning content is causally potent, challenging alignment strategies that supervise only outputs and showing reasoning carries independent signal that deeply internalizes in models.

Abstract: Chain-of-Thought (CoT) is often viewed as a window into LLM decision-making, yet recent work suggests it may function merely as post-hoc rationalization. This raises a critical alignment question: Does the reasoning trace causally shape model generalization independent of the final answer? To isolate reasoning’s causal effect, we design a controlled experiment holding final harmful answers constant while varying reasoning paths. We construct datasets with \textit{Evil} reasoning embracing malice, \textit{Misleading} reasoning rationalizing harm, and \textit{Submissive} reasoning yielding to pressure. We train models (0.6B–14B parameters) under multiple paradigms, including question-thinking-answer (QTA), question-thinking (QT), and thinking-only (T-only), and evaluate them in both think and no-think modes. We find that: (1) CoT training could amplify harmful generalization more than standard fine-tuning; (2) distinct reasoning types induce distinct behavioral patterns aligned with their semantics, despite identical final answers; (3) training on reasoning without answer supervision (QT or T-only) is sufficient to alter behavior, proving reasoning carries an independent signal; and (4) these effects persist even when generating answers without reasoning, indicating deep internalization. Our findings demonstrate that reasoning content is causally potent, challenging alignment strategies that supervise only outputs.

Method: Created a 12,000-pair dataset of matched affirmative/negated sentences. Used Negation Effect Score (NES) to measure sensitivity. Conducted activation patching (inserting affirmative activations into negated sentences) and ablation (disabling specific attention heads) to probe causal structure.

Result: Negation processing is highly concentrated in limited mid-layer attention heads (layers 4-6). Ablating these heads disrupts negation sensitivity. Activation patching shows these heads carry affirmative signals rather than restoring baseline behavior. Patterns are consistent across negation forms and detectable on external xNot360 benchmark.

Conclusion: Negation capability in GPT-2 is localized to specific components rather than distributed, providing insights into how language models process linguistic transformations and suggesting targeted interventions could improve negation handling.

Abstract: Negation remains a persistent challenge for modern language models, often causing reversed meanings or factual errors. In this work, we conduct a causal analysis of how GPT-2 Small internally processes such linguistic transformations. We examine its hidden representations at both the layer and head level. Our analysis is based on a self-curated 12,000-pair dataset of matched affirmative and negated sentences, covering multiple linguistic templates and forms of negation. To quantify this behavior, we define a metric, the Negation Effect Score (NES), which measures the model’s sensitivity in distinguishing between affirmative statements and their negations. We carried out two key interventions to probe causal structure. In activation patching, internal activations from affirmative sentences were inserted into their negated counterparts to see how meaning shifted. In ablation, specific attention heads were temporarily disabled to observe how logical polarity changed. Together, these steps revealed how negation signals move and evolve through GPT-2’s layers. Our findings indicate that this capability is not widespread; instead, it is highly concentrated within a limited number of mid-layer attention heads, primarily within layers 4 to 6. Ablating these specific components directly disrupts the model’s negation sensitivity: on our in-domain, ablation increased NES (indicating weaker negation sensitivity), and re-introducing cached affirmative activations (rescue) increased NES further, confirming that these heads carry affirmative signal rather than restoring baseline behavior. On xNot360, ablation slightly decreased NES and rescue restored performance above baseline. This pattern demonstrates that these causal patterns are consistent across various negation forms and remain detectable on the external xNot360 benchmark, though with smaller magnitude.

Method: Proposes a heterogeneous dual LLM ensemble using self-consistency and weighted voting, with a novel Deliberative Complexity Gating mechanism that uses cross-model behavioral signals and LLM response-length as a proxy for sample ambiguity. Also evaluates multi-agent debate as an alternative strategy.

Result: Achieved a Macro-F1 score of 0.85 on the evaluation set, securing 3rd place in the competition.

Conclusion: The proposed system with heterogeneous LLM ensemble and DCG mechanism effectively addresses response clarity classification, with DCG outperforming multi-agent debate by using cross-model behavioral signals rather than just increasing agent count.

Abstract: This paper describes our system for SemEval-2026 Task 6, which classifies clarity of responses in political interviews into three categories: Clear Reply, Ambivalent, and Clear Non-Reply. We propose a heterogeneous dual large language model (LLM) ensemble via self-consistency (SC) and weighted voting, and a novel post-hoc correction mechanism, Deliberative Complexity Gating (DCG). This mechanism uses cross-model behavioral signals and exploits the finding that an LLM response-length proxy correlates strongly with sample ambiguity. To further examine mechanisms for improving ambiguity detection, we evaluated multi-agent debate as an alternative strategy for increasing deliberative capacity. Unlike DCG, which adaptively gates reasoning using cross-model behavioral signals, debate increases agent count without increasing model diversity. Our solution achieved a Macro-F1 score of 0.85 on the evaluation set, securing 3rd place.

Method: Introduces ShatterMed-QA benchmark with topology-regularized medical Knowledge Graph using novel k-Shattering algorithm that prunes generic hubs to sever logical shortcuts. Uses implicit bridge entity masking and topology-driven hard negative sampling to force models to navigate biologically plausible distractors.

Result: Comprehensive evaluations of 21 LLMs show massive performance degradation on multi-hop tasks, especially among domain-specific models. Restoring masked evidence via RAG triggers near-universal performance recovery, validating the benchmark’s structural fidelity.

Conclusion: ShatterMed-QA effectively diagnoses fundamental reasoning deficits in current medical AI and demonstrates that shortcut learning is a critical barrier to authentic clinical reasoning in LLMs.

Abstract: While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is “shortcut learning”, where models exploit highly connected, generic hub nodes (e.g., “inflammation”) in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA’s structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: https://shattermed-qa-web.vercel.app/

Method: Introduces LESS (Large Language Model Enhanced Semi-supervised Learning) framework that uses LLMs to correct pseudo-labels generated on in-the-wild data. Pseudo-labeled text from ASR/AST is refined by an LLM, with further improvement through data filtering strategies.

Result: Achieved absolute Word Error Rate reduction of 3.8% on WenetSpeech (Mandarin ASR), and BLEU score increases of 0.8 and 0.7, reaching 34.0 on Callhome and 64.7 on Fisher testsets for Spanish-to-English AST.

Conclusion: LESS demonstrates effectiveness across diverse languages, tasks, and domains for semi-supervised learning on in-the-wild speech data, with recipe released as open source for further research.

Abstract: Although state-of-the-art Speech Foundational Models can produce high-quality text pseudo-labels, applying Semi-Supervised Learning (SSL) for in-the-wild real-world data remains challenging due to its richer and more complex acoustics compared to curated datasets. To address the challenges, we introduce LESS (Large Language Model Enhanced Semi-supervised Learning), a versatile framework that uses Large Language Models (LLMs) to correct pseudo-labels generated on in-the-wild data. In the LESS framework, pseudo-labeled text from Automatic Speech Recognition (ASR) or Automatic Speech Translation (AST) of the unsupervised data is refined by an LLM, and further improved by a data filtering strategy. Across Mandarin ASR and Spanish-to-English AST evaluations, LESS delivers consistent gains, with an absolute Word Error Rate reduction of 3.8% on WenetSpeech, and BLEU score increase of 0.8 and 0.7, achieving 34.0 on Callhome and 64.7 on Fisher testsets respectively. These results highlight LESS’s effectiveness across diverse languages, tasks, and domains. We have released the recipe as open source to facilitate further research in this area.

Method: Used 600 eighth-grade persuasive essays from PERSUADE dataset, generated feedback with four LLMs (GPT-4o, GPT-3.5-turbo, Llama-3.3 70B, Llama-3.1 8B) under prompt conditions embedding gender, race/ethnicity, learning needs, achievement, and motivation. Analyzed lexical shifts using adapted Marked Words framework.

Result: Revealed systematic, stereotype-aligned shifts in feedback conditioned on student attributes even with identical essay content. Feedback for students marked by race, language, or disability showed positive feedback bias and feedback withholding bias—overuse of praise, less substantive critique, and assumptions of limited ability.

Conclusion: LLMs exhibit “marked pedagogies” that tailor feedback based on presumed student attributes, reproducing social stereotypes. Highlights need for transparency and accountability in automated feedback tools to prevent biased educational outcomes.

Abstract: Effective personalized feedback is critical to students’ literacy development. Though LLM-powered tools now promise to automate such feedback at scale, LLMs are not language-neutral: they privilege standard academic English and reproduce social stereotypes, raising concerns about how “personalization” shapes the feedback students receive. We examine how four widely used LLMs (GPT-4o, GPT-3.5-turbo, Llama-3.3 70B, Llama-3.1 8B) adapt written feedback in response to student attributes. Using 600 eighth-grade persuasive essays from the PERSUADE dataset, we generated feedback under prompt conditions embedding gender, race/ethnicity, learning needs, achievement, and motivation. We analyze lexical shifts across model outputs by adapting the Marked Words framework. Our results reveal systematic, stereotype-aligned shifts in feedback conditioned on presumed student attributes–even when essay content was identical. Feedback for students marked by race, language, or disability often exhibited positive feedback bias and feedback withholding bias–overuse of praise, less substantive critique, and assumptions of limited ability. Across attributes, models tailored not only what content was emphasized but also how writing was judged and how students were addressed. We term these instructional orientations Marked Pedagogies and highlight the need for transparency and accountability in automated feedback tools.

Method: Developed an encoder-adapter-LLM architecture that combines speech encoders with text-based large language models through domain-specific adapters, using a multi-stage training strategy with extensive multilingual audio datasets.

Result: Achieved competitive Word Error Rate (WER) performance on both development and test sets, obtaining second place in the challenge ranking.

Conclusion: The encoder-adapter-LLM framework effectively leverages LLM reasoning capabilities for multilingual speech recognition, demonstrating strong performance in conversational scenarios through careful architectural design and training strategies.

Abstract: This paper describes our Triple X speech recognition system submitted to Task 1 of the Multi-Lingual Conversational Speech Language Modeling (MLC-SLM) Challenge. Our work focuses on optimizing speech recognition accuracy in multilingual conversational scenarios through an innovative encoder-adapter-LLM architecture. This framework harnesses the powerful reasoning capabilities of text-based large language models while incorporating domain-specific adaptations. To further enhance multilingual recognition performance, we adopted a meticulously designed multi-stage training strategy leveraging extensive multilingual audio datasets. Experimental results demonstrate that our approach achieves competitive Word Error Rate (WER) performance on both dev and test sets, obtaining second place in the challenge ranking.

Method: A two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. Built on Next.js with React, integrates Hugging Face inference endpoints for bias detection, and uses Vercel AI SDK for multi-provider LLM access.

Result: The system provides real-time streaming responses, per-model bias summaries, comparison views highlighting differences in bias distributions, and supports export to JSON/PDF with interactive visualizations (bar charts, radar charts).

Conclusion: LLM BiasScope is available as an open-source web application that provides a practical tool for bias evaluation and comparative analysis of LLM behavior across multiple providers.

Abstract: As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on Next.js with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.

Method: Introduced a paired-trajectory protocol replaying real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier), decomposing divergence into information-channel and memory-channel mechanisms, and testing narrative-only corruption scenarios.

Result: Consistent evaluation-blindness pattern: recommendation quality preserved under contamination (utility preservation ratio ~1.0) while risk-inappropriate products appear in 65-93% of turns. Safety violations are information-channel-driven, emerge at first contaminated turn, persist without self-correction, and narrative-only corruption induces significant drift while evading consistency monitors.

Conclusion: Standard evaluation metrics like NDCG fail to capture safety risks in multi-turn LLM agents. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, highlighting the need for trajectory-level safety monitoring beyond single-turn quality metrics for deployed agents in high-stakes settings.

Abstract: Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.

Method: Introduces Long-horizon Memory Embedding Benchmark (LMEB) with 22 datasets and 193 zero-shot retrieval tasks across 4 memory types: episodic, dialogue, semantic, and procedural, using both AI-generated and human-annotated data. Evaluates 15 embedding models ranging from hundreds of millions to ten billion parameters.

Result: (1) LMEB provides reasonable difficulty level; (2) Larger models don’t always perform better; (3) LMEB and MTEB (traditional passage retrieval benchmark) exhibit orthogonality, suggesting performance doesn’t generalize between task types.

Conclusion: No universal model excels across all memory retrieval tasks, and traditional passage retrieval performance doesn’t generalize to long-horizon memory retrieval. LMEB fills a crucial gap in memory embedding evaluation and drives advancements in text embedding for long-term, context-dependent memory retrieval.

Abstract: Memory embeddings are crucial for memory-augmented systems, such as OpenClaw, but their evaluation is underexplored in current text embedding benchmarks, which narrowly focus on traditional passage retrieval and fail to assess models’ ability to handle long-horizon memory retrieval tasks involving fragmented, context-dependent, and temporally distant information. To address this, we introduce the Long-horizon Memory Embedding Benchmark (LMEB), a comprehensive framework that evaluates embedding models’ capabilities in handling complex, long-horizon memory retrieval tasks. LMEB spans 22 datasets and 193 zero-shot retrieval tasks across 4 memory types: episodic, dialogue, semantic, and procedural, with both AI-generated and human-annotated data. These memory types differ in terms of level of abstraction and temporal dependency, capturing distinct aspects of memory retrieval that reflect the diverse challenges of the real world. We evaluate 15 widely used embedding models, ranging from hundreds of millions to ten billion parameters. The results reveal that (1) LMEB provides a reasonable level of difficulty; (2) Larger models do not always perform better; (3) LMEB and MTEB exhibit orthogonality. This suggests that the field has yet to converge on a universal model capable of excelling across all memory retrieval tasks, and that performance in traditional passage retrieval may not generalize to long-horizon memory retrieval. In summary, by providing a standardized and reproducible evaluation framework, LMEB fills a crucial gap in memory embedding evaluation, driving further advancements in text embedding for handling long-term, context-dependent memory retrieval. LMEB is available at https://github.com/KaLM-Embedding/LMEB.

Method: EPT decomposes task adaptation into two stages: (1) a shared meta-knowledge subspace encoding universal linguistic patterns in low dimensions, and (2) a pyramid projection mechanism using learnable up-projection operators to reconstruct high-dimensional features at varying scales, with a task-aware router dynamically selecting optimal multi-scale feature combinations.

Result: Extensive experiments across multiple multi-task benchmarks show EPT significantly outperforms state-of-the-art MoE-LoRA variants while simultaneously reducing the number of training parameters through re-parameterization.

Conclusion: EPT successfully bridges the gap in existing PEFT methods by incorporating multi-scale feature pyramids, enabling more effective adaptation to diverse task complexities while maintaining parameter efficiency.

Abstract: Parameter-Efficient Fine-Tuning (PEFT) has become a dominant paradigm for deploying LLMs in multi-task scenarios due to its extreme parameter efficiency. While Mixture-of-Experts (MoE) based LoRA variants have achieved promising results by dynamically routing tokens to different low-rank experts, they largely overlook the hierarchical nature of task complexity. Existing methods typically employ experts with uniform architectures, limiting their ability to capture diverse feature granularities required by distinct tasks–where some tasks demand high-level semantic abstraction while others require fine-grained syntactic manipulation. To bridge this gap, we propose Expert Pyramid Tuning (EPT), a novel architecture that integrates the multi-scale feature pyramid concept from computer vision into the realm of PEFT. Unlike standard LoRA, EPT decomposes task adaptation into two stages: (1) A shared meta-knowledge Subspace that encodes universal linguistic patterns in low dimensions; (2) A Pyramid Projection Mechanism that utilizes learnable up-projection operators to reconstruct high-dimensional features at varying scales. A task-aware router then dynamically selects the optimal combination of these multi-scale features. Extensive experiments across multiple multi-task benchmarks demonstrate that EPT significantly outperforms SOTA MoE-LoRA variants. Crucially, thanks to the re-parameterization capability of our design, EPT achieves this performance improvement while simultaneously reducing the number of training parameters.

Method: Uses off-the-shelf RTD discriminator to localize suspicious tokens (similar to word-substitution perturbations), masks them, and detects adversarial examples by observing prediction confidence shift of victim model before/after intervention with only two black-box queries.

Result: Effectively detects adversarial texts from diverse state-of-the-art attack methods across multiple benchmark datasets, surpassing existing detection baselines across multiple metrics.

Conclusion: RTD-Guard offers highly efficient, practical, resource-light defense mechanism particularly suited for real-world deployment in resource-constrained or privacy-sensitive environments.

Abstract: Textual adversarial attacks pose a serious security threat to Natural Language Processing (NLP) systems by introducing imperceptible perturbations that mislead deep learning models. While adversarial example detection offers a lightweight alternative to robust training, existing methods typically rely on prior knowledge of attacks, white-box access to the victim model, or numerous queries, which severely limits their practical deployment. This paper introduces RTD-Guard, a novel black-box framework for detecting textual adversarial examples. Our key insight is that word-substitution perturbations in adversarial attacks closely resemble the “replaced tokens” that a Replaced Token Detection (RTD) discriminator is pre-trained to identify. Leveraging this, RTD-Guard employs an off-the-shelf RTD discriminator-without fine-tuning-to localize suspicious tokens, masks them, and detects adversarial examples by observing the prediction confidence shift of the victim model before and after intervention. The entire process requires no adversarial data, model tuning, or internal model access, and uses only two black-box queries. Comprehensive experiments on multiple benchmark datasets demonstrate that RTD-Guard effectively detects adversarial texts generated by diverse state-of-the-art attack methods. It surpasses existing detection baselines across multiple metrics, offering a highly efficient, practical, and resource-light defense mechanism-particularly suited for real-world deployment in resource-constrained or privacy-sensitive environments.

Method: Designed around Human-AI teaming principles with workflows for verifying and curating data, featuring iterative review/correction of AI outputs and using this interaction as feedback to improve future LLM-based inference.

Result: Evaluation through intrinsic benchmarking and real-world case studies across multiple domains shows SCILIRE improves extraction fidelity and facilitates efficient dataset creation.

Conclusion: SCILIRE effectively addresses the challenge of extracting structured knowledge from scientific literature through Human-AI collaboration, improving both accuracy and efficiency.

Abstract: The rapid growth of scientific literature has made manual extraction of structured knowledge increasingly impractical. To address this challenge, we introduce SCILIRE, a system for creating datasets from scientific literature. SCILIRE has been designed around Human-AI teaming principles centred on workflows for verifying and curating data. It facilitates an iterative workflow in which researchers can review and correct AI outputs. Furthermore, this interaction is used as a feedback signal to improve future LLM-based inference. We evaluate our design using a combination of intrinsic benchmarking outcomes together with real-world case studies across multiple domains. The results demonstrate that SCILIRE improves extraction fidelity and facilitates efficient dataset creation.

Method: Three-stage optimization: 1) Custom CK Flash Attention operator for ONNX Runtime on ROCm reduces attention memory from O(n²) to O(n); 2) NLP prompt compression techniques (TextRank, TF-IDF, etc.) reduce inputs to ~512 tokens; 3) Near-streaming body processing with adaptive chunking and zero-copy JSON.

Result: 98× improvement (4,918ms to 50ms), 16K-token routing in 108ms, total router GPU footprint under 800MB, enabling router to share GPU with LLM serving without dedicated accelerator.

Conclusion: The optimized vLLM semantic router solves both latency and memory problems for long-context classification, making it practical to co-locate routers with LLM serving instances on shared GPUs.

Abstract: System-level routers that intercept LLM requests for safety classification, domain routing, and PII detection must be both fast and operationally lightweight: they should add minimal latency to every request, yet not require a dedicated GPU – an expensive resource better used for LLM inference itself. When the router co-locates on the same GPU as vLLM serving instances, standard attention’s $O(n^2)$ memory makes long-context classification (8K–32K tokens) impossible: at 8K tokens, three concurrent classifiers need ${\sim}$4.5,GB for attention masks alone, far exceeding the memory left by vLLM. We present three staged optimizations for the vLLM Semantic Router, benchmarked on AMD Instinct MI300X, that solve both the latency and the memory problem. \emph{Stage1}: a custom CK Flash Attention operator for ONNX Runtime on ROCm reduces attention memory from $O(n^2)$ to $O(n)$ and end-to-end (E2E) latency from 4{,}918,ms to 127,ms (\textbf{38.7$\times$}), enabling 8K–32K tokens where SDPA OOMs. \emph{Stage2}: classical NLP prompt compression (TextRank, position weighting, TF-IDF, and novelty scoring) reduces all inputs to ${\sim}$512 tokens without neural inference, capping both latency and GPU memory at a constant regardless of original prompt length (E2E 127$\to$62,ms, \textbf{2.0$\times$}). \emph{Stage3}: near-streaming body processing with adaptive chunking and zero-copy JSON eliminates serialization overhead (E2E 62$\to$50,ms, \textbf{1.2$\times$}). Cumulatively: \textbf{98$\times$} improvement (4{,}918,ms to 50,ms), 16K-token routing in 108,ms, and a total router GPU footprint under 800,MB – small enough to share a GPU with LLM serving and removing the need for a dedicated accelerator. Stage1 targets AMD ROCm (NVIDIA GPUs already have FlashAttention via cuDNN); Stages2 and3 are hardware-agnostic.

Method: Systematic survey structured around three training stages: continual pre-training, continual fine-tuning, and continual alignment. Taxonomy includes rehearsal-, regularization-, and architecture-based methods with further subdivision by forgetting mitigation mechanisms.

Result: Provides comparative analysis of adaptability and improvements of traditional CL methods for LLMs, highlighting core distinctions between LLM CL and traditional ML regarding scale, parameter efficiency, and emergent capabilities.

Conclusion: Current methods show promise in specific domains but fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. The survey provides a structured framework for understanding current achievements and future opportunities in lifelong learning for LLMs.

Abstract: Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual alignment.Beyond the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.

Method: Proposes TESS with a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. Uses LLMs with structured prompting to extract interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) from text, filtered through confidence-aware gating.

Result: Experiments on four real-world datasets show up to 29% reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines.

Conclusion: TESS effectively bridges the modality gap between text and time-series data by creating an interpretable intermediate representation that enables better multimodal fusion for forecasting.

Abstract: Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.

Method: Proposes MetaKE, a framework that treats knowledge editing as bi-level optimization: upper-level optimizer learns feasible edit targets to maximize post-edit performance, lower-level solver executes editing; uses Structural Gradient Proxy to backpropagate editability constraints through complex solvers

Result: MetaKE significantly outperforms strong baselines in knowledge editing tasks, demonstrating improved alignment between semantic targets and model’s feasible manifold

Conclusion: MetaKE offers a new perspective on knowledge editing by addressing the semantic-execution disconnect through bi-level optimization and meta-learning aligned targets

Abstract: Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical “Semantic-Execution Disconnect”: the semantic target is derived independently without feedback from the downstream’s feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model’s feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.

Method: Used text similarity metrics to evaluate different ChatGPT models’ capacity to generate diverse textual outputs, testing across multiple model versions while explicitly setting temperature parameter to one to encourage diversity.

Result: Found measurable decline in recent ChatGPT releases’ ability to produce varied text, with outputs becoming increasingly similar across different versions - a phenomenon termed “model self-convergence.”

Conclusion: The reduction in output diversity is likely due to synthetic data infiltration in training datasets, suggesting that recursive training on LLM-generated data leads to model degradation over time through self-convergence.

Abstract: Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models’ capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases’ ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.

Method: Proposes a solution-conditioned adversarial verification framework that iteratively refines test cases based on candidate solution execution behaviors, creating EvolveCoder-22k dataset through multiple rounds of adversarial test case evolution.

Result: Iterative refinement substantially strengthens verification (pass@1 decreased from 43.80 to 31.22). RL on EvolveCoder-22k improves Qwen3-4B by average 4.2 points across four benchmarks, outperforming strong 4B-scale baselines.

Conclusion: Adversarial, solution-conditioned verification is crucial for effective and scalable reinforcement learning in code generation, as demonstrated by the success of EvolveCoder-22k dataset.

Abstract: Reinforcement learning with verifiable rewards (RLVR) is a promising approach for improving code generation in large language models, but its effectiveness is limited by weak and static verification signals in existing coding RL datasets. In this paper, we propose a solution-conditioned and adversarial verification framework that iteratively refines test cases based on the execution behaviors of candidate solutions, with the goal of increasing difficulty, improving discriminative power, and reducing redundancy. Based on this framework, we introduce EvolveCoder-22k, a large-scale coding reinforcement learning dataset constructed through multiple rounds of adversarial test case evolution. Empirical analysis shows that iterative refinement substantially strengthens verification, with pass@1 decreasing from 43.80 to 31.22. Reinforcement learning on EvolveCoder-22k yields stable optimization and consistent performance gains, improving Qwen3-4B by an average of 4.2 points across four downstream benchmarks and outperforming strong 4B-scale baselines. Our results highlight the importance of adversarial, solution-conditioned verification for effective and scalable reinforcement learning in code generation.

Method: Learns construction grammars from utterances annotated with constituency structure and semantic frames, formalized within the Fluid Construction Grammar framework, resulting in networks of tens of thousands of constructions.

Result: Produces human-interpretable computational construction grammars that support frame-semantic analysis of open-domain text and contain extensive information about syntactico-semantic usage patterns.

Conclusion: The method contributes to scaling constructionist approaches, corroborates scalability of fundamental construction grammar conjectures, and provides practical tools for studying English argument structure in broad-coverage corpora.

Abstract: We present a method for learning large-scale, broad-coverage construction grammars from corpora of language use. Starting from utterances annotated with constituency structure and semantic frames, the method facilitates the learning of human-interpretable computational construction grammars that capture the intricate relationship between syntactic structures and the semantic relations they express. The resulting grammars consist of networks of tens of thousands of constructions formalised within the Fluid Construction Grammar framework. Not only do these grammars support the frame-semantic analysis of open-domain text, they also house a trove of information about the syntactico-semantic usage patterns present in the data they were learnt from. The method and learnt grammars contribute to the scaling of usage-based, constructionist approaches to language, as they corroborate the scalability of a number of fundamental construction grammar conjectures while also providing a practical instrument for the constructionist study of English argument structure in broad-coverage corpora.

Method: Created SectEval benchmark with 88 questions in English and Hindi to test 15 LLMs (proprietary and open-weights). Evaluated model responses for bias patterns across languages and tested location-based adaptation by simulating users from different countries.

Result: Major language inconsistency: powerful models like DeepSeek-v3 and GPT-4o favored Shia answers in English but Sunni answers in Hindi. Location-aware models like Claude-3.5 adapted answers to match user’s country (Shia for Iran, Sunni for Saudi Arabia), while smaller Hindi models consistently favored Sunni regardless of location.

Conclusion: AI religious knowledge is not neutral; LLM responses vary significantly based on language and claimed location, showing systematic biases that could provide different religious advice to users based on these factors.

Abstract: As Large Language Models (LLMs) becomes a popular source for religious knowledge, it is important to know if it treats different groups fairly. This study is the first to measure how LLMs handle the differences between the two main sects of Islam: Sunni and Shia. We present a test called SectEval, available in both English and Hindi, consisting of 88 questions, to check the bias-ness of 15 top LLM models, both proprietary and open-weights. Our results show a major inconsistency based on language. In English, many powerful models DeepSeek-v3 and GPT-4o often favored Shia answers. However, when asked the exact same questions in Hindi, these models switched to favoring Sunni answers. This means a user could get completely different religious advice just by changing languages. We also looked at how models react to location. Advanced models Claude-3.5 changed their answers to match the user’s country-giving Shia answers to a user from Iran and Sunni answers to a user from Saudi Arabia. In contrast, smaller models (especially in Hindi) ignored the user’s location and stuck to a Sunni viewpoint. These findings show that AI is not neutral; its religious ``truth’’ changes depending on the language you speak and the country you claim to be from. The data set is available at https://github.com/secteval/SectEval/

Method: SteerRM uses Sparse Autoencoder (SAE)-based interventions: isolates stylistic effects using contrastive paired responses, identifies bias-related SAE features with a strength-stability criterion, and suppresses them at inference time.

Result: Across six reward models on RM-Bench, SteerRM improves Hard-split accuracy by 7.3 points on average while preserving overall performance. Results generalize across RM architectures and bias types, with format-related features concentrated in shallow layers and transferable across models.

Conclusion: SAE-based interventions can effectively mitigate reward-model biases without retraining, providing a practical and interpretable solution for alignment pipelines, revealing shared architecture-level bias encoding patterns.

Abstract: Reward models (RMs) are critical components of alignment pipelines, yet they exhibit biases toward superficial stylistic cues, preferring better-presented responses over semantically superior ones. Existing debiasing methods typically require retraining or architectural modifications, while direct activation suppression degrades performance due to representation entanglement. We propose SteerRM, the first training-free method for debiasing reward models using Sparse Autoencoder (SAE)-based interventions. SteerRM isolates stylistic effects using contrastive paired responses, identifies bias-related SAE features with a strength-stability criterion, and suppresses them at inference time. Across six reward models on RM-Bench, SteerRM improves Hard-split accuracy by 7.3 points on average while preserving overall performance. Results on a Gemma-based reward model and a controlled non-format bias further suggest generalization across RM architectures and bias types. We further find that format-related features are concentrated in shallow layers and transfer across models, revealing shared architecture-level bias encoding patterns. These results show that SAE-based interventions can mitigate reward-model biases without retraining, providing a practical and interpretable solution for alignment pipelines.

Method: AVR inserts a lightweight semantic routing layer that estimates action difficulty from multimodal embeddings, probes small VLMs for confidence, and routes actions to the cheapest model meeting target reliability thresholds, with context retrieval for warm agents.

Result: AVR achieves up to 78% inference cost reduction while staying within 2 percentage points of all-large-model baseline accuracy, and integrates safety by escalating high-risk actions to strongest models.

Conclusion: AVR provides an efficient framework for model routing in GUI automation that balances cost and accuracy while maintaining safety through intelligent escalation of difficult actions.

Abstract: Computer Use Agents (CUAs) translate natural-language instructions into Graphical User Interface (GUI) actions such as clicks, keystrokes, and scrolls by relying on a Vision-Language Model (VLM) to interpret screenshots and predict grounded tool calls. However, grounding accuracy varies dramatically across VLMs, while current CUA systems typically route every action to a single fixed model regardless of difficulty. We propose \textbf{Adaptive VLM Routing} (AVR), a framework that inserts a lightweight semantic routing layer between the CUA orchestrator and a pool of VLMs. For each tool call, AVR estimates action difficulty from multimodal embeddings, probes a small VLM to measure confidence, and routes the action to the cheapest model whose predicted accuracy satisfies a target reliability threshold. For \textit{warm} agents with memory of prior UI interactions, retrieved context further narrows the capability gap between small and large models, allowing many actions to be handled without escalation. We formalize routing as a cost–accuracy trade-off, derive a threshold-based policy for model selection, and evaluate AVR using ScreenSpot-Pro grounding data together with the OpenClaw agent routing benchmark. Across these settings, AVR projects inference cost reductions of up to 78% while staying within 2 percentage points of an all-large-model baseline. When combined with the Visual Confused Deputy guardrail, AVR also escalates high-risk actions directly to the strongest available model, unifying efficiency and safety within a single routing framework. Materials are also provided Model, benchmark, and code: https://github.com/vllm-project/semantic-router.

Method: Systematic investigation of option design impact on RLVR, revealing two key insights: 1) Mismatches in option counts between training/testing degrade performance, 2) Strong distractors mitigate random guessing. Proposed Iterative Distractor Curation (IDC) framework actively constructs high-quality distractors to block elimination shortcuts.

Result: Experiments on various benchmarks show IDC effectively enhances distractor quality and yields significant gains in RLVR training compared to original data.

Conclusion: Option design significantly impacts RLVR effectiveness; strong distractors enable effective training even with 2-way questions; IDC framework successfully improves reasoning by blocking shortcuts through active distractor curation.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capabilities of Large Language Models. When applied to RLVR, Multiple-Choice Questions (MCQs) offer a scalable source of verifiable data but risk inducing reward hacking, where models shortcut reasoning via random guessing or simple elimination. Current approaches often mitigate this by converting MCQs to open-ended formats, thereby discarding the contrastive signal provided by expert-designed distractors. In this work, we systematically investigate the impact of option design on RLVR. Our analysis highlights two primary insights: (1) Mismatches in option counts between training and testing degrade performance. (2) Strong distractors effectively mitigate random guessing, enabling effective RLVR training even with 2-way questions. Motivated by these findings, we propose Iterative Distractor Curation (IDC), a framework that actively constructs high-quality distractors to block elimination shortcuts and promote deep reasoning. Experiments on various benchmarks demonstrate that our method effectively enhances distractor quality and yields significant gains in RLVR training compared to the original data.

Method: Created a leaderboard with associated benchmarks specifically designed for European Portuguese, including novel benchmarks for model safeguards and alignment to Portuguese culture.

Result: Established a publicly available leaderboard at https://huggingface.co/spaces/PORTULAN/portuguese-llm-leaderboard with comprehensive benchmarks for European Portuguese LLMs.

Conclusion: The leaderboard fills an important gap in evaluating LLMs for European Portuguese and provides novel benchmarks addressing safety and cultural alignment aspects.

Abstract: This paper reports on the development of a leaderboard of Open Large Language Models (LLM) for European Portuguese (PT-PT), and on its associated benchmarks. This leaderboard comes as a way to address a gap in the evaluation of LLM for European Portuguese, which so far had no leaderboard dedicated to this variant of the language. The paper also reports on novel benchmarks, including some that address aspects of performance that so far have not been available in benchmarks for European Portuguese, namely model safeguards and alignment to Portuguese culture. The leaderboard is available at https://huggingface.co/spaces/PORTULAN/portuguese-llm-leaderboard.

Method: Extends BabyBERTa to English-French scenarios with strictly size-matched data conditions. Compares two training corpora types: (1) child-directed speech (~2.5M tokens) and (2) multi-domain corpora (~10M tokens). Introduces new evaluation resources including French versions of QAMR and QASRL, plus English and French multi-domain corpora. Evaluates models on syntactic and semantic tasks and compares with Wikipedia-only trained models.

Result: Training on Wikipedia consistently benefits semantic tasks, while child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT architectures, suggesting consistent trends.

Conclusion: The study reveals context-dependent effects of different training data types on model performance in multilingual settings, with specific benefits for different task types depending on training corpus. The findings provide insights into developmentally plausible language modeling across languages.

Abstract: Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.

Method: Built on pretrained multilingual BERT, integrates contextual representations with handcrafted linguistic features, jointly optimizes fine-grained multi-label abuse classification and three-class main classification tasks, and uses iterative self-training with confidence-based pseudo-labeling for cross-lingual knowledge transfer.

Result: Achieves macro F1-score of up to 0.9847 on multi-label task and accuracy of 0.6775 on main classification task across four public datasets, with ablation studies verifying component effectiveness.

Conclusion: HMS-BERT effectively addresses multilingual and multi-label cyberbullying detection challenges through hybrid multi-task learning and self-training for cross-lingual knowledge transfer.

Abstract: Cyberbullying on social media is inherently multilingual and multi-faceted, where abusive behaviors often overlap across multiple categories. Existing methods are commonly limited by monolingual assumptions or single-task formulations, which restrict their effectiveness in realistic multilingual and multi-label scenarios. In this paper, we propose HMS-BERT, a hybrid multi-task self-training framework for multilingual and multi-label cyberbullying detection. Built upon a pretrained multilingual BERT backbone, HMS-BERT integrates contextual representations with handcrafted linguistic features and jointly optimizes a fine-grained multi-label abuse classification task and a three-class main classification task. To address labeled data scarcity in low-resource languages, an iterative self-training strategy with confidence-based pseudo-labeling is introduced to facilitate cross-lingual knowledge transfer. Experiments on four public datasets demonstrate that HMS-BERT achieves strong performance, attaining a macro F1-score of up to 0.9847 on the multi-label task and an accuracy of 0.6775 on the main classification task. Ablation studies further verify the effectiveness of the proposed components.

Method: 1) Generate task-informed keywords for comprehensive domain coverage; 2) Create diverse instructions by pairing keywords with different cognitive levels from Bloom’s Taxonomy; 3) Use self-consistency validation to ensure data quality.

Result: Models fine-tuned on DS²-Instruct generated data achieve substantial improvements over existing data generation methods across seven challenging domains including mathematics, finance, and logical reasoning.

Conclusion: DS²-Instruct provides an effective zero-shot framework for generating high-quality domain-specific instruction datasets without human supervision, enabling better adaptation of LLMs to specialized domains.

Abstract: Adapting Large Language Models (LLMs) to specialized domains requires high-quality instruction tuning datasets, which are expensive to create through human annotation. Existing data synthesis methods focus on general-purpose tasks and fail to capture domain-specific terminology and reasoning patterns. To address this, we introduce DS$^2$-Instruct, a zero-shot framework that generates domain-specific instruction datasets without human supervision. Our approach first generates task-informed keywords to ensure comprehensive domain coverage. It then creates diverse instructions by pairing these keywords with different cognitive levels from Bloom’s Taxonomy. Finally, it uses self-consistency validation to ensure data quality. We apply this framework to generate datasets across seven challenging domains, such as mathematics, finance, and logical reasoning. Comprehensive evaluation demonstrates that models fine-tuned on our generated data achieve substantial improvements over existing data generation methods.

Method: Created benchmark with 5 subtasks, collected instruction/preference data through multi-stage process, tested 20+ reward models (classifiers and generative), and introduced novel Long-form Needle-in-a-Haystack Test.

Result: Current models lack long-form reward modeling capabilities; found correlation between performance and error position/response length; classifiers show better generalizability than generative models on same data.

Conclusion: First comprehensive benchmark for long-form reward modeling that reveals current limitations and provides platform for future progress in this crucial area.

Abstract: The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error’s position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifiers exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

Method: Proposes Iterative MBR Distillation framework using Minimum Bayes Risk decoding with off-the-shelf LLM to generate pseudo-labels for training ESD models without human annotations.

Result: Models trained on self-generated pseudo-labels outperform both base models and supervised baselines on human annotations at system and span levels, while maintaining competitive sentence-level performance on WMT datasets.

Conclusion: The self-evolution framework effectively eliminates reliance on expensive human annotations for Error Span Detection, achieving superior performance through automated pseudo-label generation.

Abstract: Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels.Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.

Method: PCA sweep procedure that treats dimensionality selection as joint optimization over representation capacity, gradient interpretability, and stability across nearby K values.

Result: Method yields stable, interpretable Admiration-related gradient in AI discourse analysis, contrasting optimistic vs. distrustful framings, while no robust alignment for Rivalry narcissism.

Conclusion: PCA sweep constrains researcher degrees of freedom while preserving SSD’s interpretive aims, supporting transparent and psychologically meaningful analyses of connotative meaning.

Abstract: Supervised Semantic Differential (SSD) is a mixed quantitative-interpretive method that models how text meaning varies with continuous individual-difference variables by estimating a semantic gradient in an embedding space and interpreting its poles through clustering and text retrieval. SSD applies PCA before regression, but currently no systematic method exists for choosing the number of retained components, introducing avoidable researcher degrees of freedom in the analysis pipeline. We propose a PCA sweep procedure that treats dimensionality selection as a joint criterion over representation capacity, gradient interpretability, and stability across nearby values of K. We illustrate the method on a corpus of short posts about artificial intelligence written by Prolific participants who also completed Admiration and Rivalry narcissism scales. The sweep yields a stable, interpretable Admiration-related gradient contrasting optimistic, collaborative framings of AI with distrustful and derisive discourse, while no robust alignment emerges for Rivalry. We also show that a counterfactual using a high-PCA dimension solution heuristic produces diffuse, weakly structured clusters instead, reinforcing the value of the sweep-based choice of K. The case study shows how the PCA sweep constrains researcher degrees of freedom while preserving SSD’s interpretive aims, supporting transparent and psychologically meaningful analyses of connotative meaning.

Method: Uses reinforcement learning with quality estimation models, but addresses failure modes (“holes”) in existing QE models through word alignment and language alignment techniques to create better rewards for RL training.

Result: The model trained with WALAR outperforms LLaMAX (one of the strongest open-source multilingual LLMs) by a large margin on 1400 language directions on the Flores-101 dataset.

Conclusion: WALAR successfully improves LLM translation capabilities for low-resource languages using only monolingual text, addressing the parallel data scarcity problem while maintaining high-resource language performance.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capability in machine translation on high-resource language pairs, yet their performance on low-resource translation still lags behind. Existing post-training methods rely heavily on high-quality parallel data, which are often scarce or unavailable for low-resource languages. In this paper, we introduce WALAR, a reinforcement training method using only monolingual text to elevate LLMs’ translation capabilities on massive low-resource languages while retaining their performance on high-resource languages. Our key insight is based on the observation of failure modes (or “holes”) in existing source-based multilingual quality estimation (QE) models. Reinforcement learning (RL) using these QE models tends to amplify such holes, resulting in poorer multilingual LLMs. We develop techniques including word alignment and language alignment to mitigate such holes in WALAR’s reward for RL training. We continually trained an LLM supporting translation of 101 languages using WALAR. The experiments show that our new model outperforms LLaMAX, one of the strongest open-source multilingual LLMs by a large margin on 1400 language directions on Flores-101 dataset.

Method: Created ESG-Bench dataset with human-annotated QA pairs grounded in real ESG reports, with labels indicating factual support vs. hallucinations. Used task-specific Chain-of-Thought prompting strategies and fine-tuned multiple state-of-the-art LLMs on CoT-annotated rationales.

Result: CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations. The gains transfer to existing QA benchmarks beyond the ESG domain, showing broader applicability.

Conclusion: ESG-Bench enables systematic evaluation of LLMs’ ability to extract and reason over ESG content while mitigating hallucinations in compliance-critical settings. The approach provides a framework for trustworthy analysis of complex documents.

Abstract: As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms’ long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs’ ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.

Method: NAIT evaluates IT data impact by analyzing similarity of neuron activation patterns between IT dataset and target domain capabilities. Captures neuron activation patterns from in-domain datasets to construct reusable features, then selects samples based on similarity to expected activation features of target capabilities.

Result: Training on 10% Alpaca-GPT4 IT data subset selected by NAIT outperforms methods using external advanced models or uncertainty-based features across various tasks. Reveals transferability of neuron activation features across different LLM capabilities.

Conclusion: NAIT provides an efficient framework for selecting high-quality instruction tuning data. Logical reasoning and programmatic features have strong general transferability, while a stable core subset can consistently activate fundamental model capabilities across diverse tasks.

Abstract: Instruction Tuning (IT) has been proven to be an effective approach to unlock the powerful capabilities of large language models (LLMs). Recent studies indicate that excessive IT data can degrade LLMs performance, while carefully selecting a small subset of high-quality IT data can significantly enhance their capabilities. Therefore, identifying the most efficient subset data from the IT dataset to effectively develop either specific or general abilities in LLMs has become a critical challenge. To address this, we propose a novel and efficient framework called NAIT. NAIT evaluates the impact of IT data on LLMs performance by analyzing the similarity of neuron activation patterns between the IT dataset and the target domain capability. Specifically, NAIT captures neuron activation patterns from in-domain datasets of target domain capabilities to construct reusable and transferable neuron activation features. It then evaluates and selects optimal samples based on the similarity between candidate samples and the expected activation features of the target capabilities. Experimental results show that training on the 10% Alpaca-GPT4 IT data subset selected by NAIT consistently outperforms methods that rely on external advanced models or uncertainty-based features across various tasks. Our findings also reveal the transferability of neuron activation features across different capabilities of LLMs. In particular, IT data with more logical reasoning and programmatic features possesses strong general transferability, enabling models to develop stronger capabilities across multiple tasks, while a stable core subset of data is sufficient to consistently activate fundamental model capabilities and universally improve performance across diverse tasks.

Method: Conducts a comprehensive comparative theoretical analysis between softmax attention and linear attention mechanisms, examining their mathematical properties and behaviors to understand the performance differences.

Result: The analysis reveals the underlying reasons for softmax attention’s superior performance over linear attention, explaining why softmax attention captures token interactions more effectively despite its higher computational cost.

Conclusion: The paper provides theoretical insights into the performance gap between softmax and linear attention, helping researchers understand the trade-offs between computational efficiency and model effectiveness in transformer architectures.

Abstract: Large transformer models have achieved state-of-the-art results in numerous natural language processing tasks. Among the pivotal components of the transformer architecture, the attention mechanism plays a crucial role in capturing token interactions within sequences through the utilization of softmax function. Conversely, linear attention presents a more computationally efficient alternative by approximating the softmax operation with linear complexity. However, it exhibits substantial performance degradation when compared to the traditional softmax attention mechanism. In this paper, we bridge the gap in our theoretical understanding of the reasons behind the practical performance gap between softmax and linear attention. By conducting a comprehensive comparative analysis of these two attention mechanisms, we shed light on the underlying reasons for why softmax attention outperforms linear attention in most scenarios.

Method: Modify loss function of pre-trained multimodal models by restricting top K softmax outputs during fine-tuning to improve adversarial robustness against popular attacks.

Result: Experiments show that after fine-tuning with the modified loss function, adversarial robustness of pre-trained multimodal models can be significantly improved against popular attacks.

Conclusion: The approach effectively improves multimodal adversarial robustness, but further research is needed on output diversity, generalization, and robustness-performance trade-offs of such loss functions.

Abstract: As Large Language Models make a breakthrough in natural language processing tasks (NLP), multimodal technique becomes extremely popular. However, it has been shown that multimodal NLP are vulnerable to adversarial attacks, where the outputs of a model can be dramatically changed by a perturbation to the input. While several defense techniques have been proposed both in computer vision and NLP models, the multimodal robustness of models have not been fully explored. In this paper, we study the adversarial robustness provided by modifying loss function of pre-trained multimodal models, by restricting top K softmax outputs. Based on the evaluation and scoring, our experiments show that after a fine-tuning, adversarial robustness of pre-trained models can be significantly improved, against popular attacks. Further research should be studying, such as output diversity, generalization and the robustness-performance trade-off of this kind of loss functions. Our code will be available after this paper is accepted

Method: Used natural language processing and machine learning (Multinomial Naive Bayes classifier) to analyze over 2000 Flamenco lyrics, categorize them into palos (genres), identify semantic fields through automatic word usage analysis, and perform network analysis using inter-genre distance metrics.

Result: Lexical variation enables accurate palo identification; semantic fields characterize each style; network analysis reveals historical connections and palo evolutions.

Conclusion: The work illuminates intricate relationships and cultural significance in Flamenco lyrics, complementing qualitative discussions with quantitative analysis and sparking new discussions on traditional music genre origins.

Abstract: Flamenco, recognized by UNESCO as part of the Intangible Cultural Heritage of Humanity, is a profound expression of cultural identity rooted in Andalusia, Spain. However, there is a lack of quantitative studies that help identify characteristic patterns in this long-lived music tradition. In this work, we present a computational analysis of Flamenco lyrics, employing natural language processing and machine learning to categorize over 2000 lyrics into their respective Flamenco genres, termed as $\textit{palos}$. Using a Multinomial Naive Bayes classifier, we find that lexical variation across styles enables to accurately identify distinct $\textit{palos}$. More importantly, from an automatic method of word usage, we obtain the semantic fields that characterize each style. Further, applying a metric that quantifies the inter-genre distance we perform a network analysis that sheds light on the relationship between Flamenco styles. Remarkably, our results suggest historical connections and $\textit{palo}$ evolutions. Overall, our work illuminates the intricate relationships and cultural significance embedded within Flamenco lyrics, complementing previous qualitative discussions with quantitative analyses and sparking new discussions on the origin and development of traditional music genres.

Method: Proposes the Superficial Safety Alignment Hypothesis (SSAH), identifying four types of attribute-critical components at neuron level: Safety Critical Units (SCU), Utility Critical Units (UCU), Complex Units (CU), and Redundant Units (RU). Uses component freezing and leveraging redundant units as “alignment budget” during fine-tuning.

Result: Shows that freezing safety-critical components allows models to retain safety while adapting to new tasks. Leveraging redundant units minimizes alignment tax while achieving alignment goals. Demonstrates safety alignment can be simplified to neuron-level functional units.

Conclusion: Safety alignment in LLMs operates at the neuron level and doesn’t need to be complicated. A few essential components can establish effective safety guardrails while maintaining model utility and minimizing alignment tax.

Abstract: As large language models (LLMs) are overwhelmingly more and more integrated into various applications, ensuring they generate safe responses is a pressing need. Previous studies on alignment have largely focused on general instruction-following but have often overlooked the distinct properties of safety alignment, such as the brittleness of safety mechanisms. To bridge the gap, we propose the Superficial Safety Alignment Hypothesis (SSAH), which posits that safety alignment teaches an otherwise unsafe model to choose the correct reasoning direction-fulfill or refuse users’ requests-interpreted as an implicit binary classification task. Through SSAH, we hypothesize that only a few essential components can establish safety guardrails in LLMs. We successfully identify four types of attribute-critical components: Safety Critical Unit (SCU), Utility Critical Unit (UCU), Complex Unit (CU), and Redundant Unit (RU). Our findings show that freezing certain safety-critical components during fine-tuning allows the model to retain its safety attributes while adapting to new tasks. Similarly, we show that leveraging redundant units in the pre-trained model as an “alignment budget” can effectively minimize the alignment tax while achieving the alignment goal. All considered, this paper concludes that the atomic functional unit for safety in LLMs is at the neuron level and underscores that safety alignment should not be complicated. We have code implementation and other information on the project website: https://ssa-h.github.io/.

Method: MEAP randomly masks a small fraction of input tokens and performs standard next-token prediction autoregressively using decoder-only Transformers, integrating MLM into NTP without bidirectional attention or encoder-decoder architectures.

Result: MEAP substantially outperforms NTP on key information retrieval and long-context reasoning tasks, performs on par or better on commonsense reasoning, and shows 11.77 percentage point advantage in lost-in-the-middle scenarios during supervised fine-tuning.

Conclusion: MEAP is a promising training paradigm that enhances LLMs’ retrieval capabilities by promoting more distinguishable attention scores and focusing on task-relevant signals while mitigating peripheral context influence.

Abstract: Large Language Models (LLMs) are discovered to suffer from accurately retrieving key information. To address this, we propose Mask-Enhanced Autoregressive Prediction (MEAP), a simple yet effective training paradigm that seamlessly integrates Masked Language Modeling (MLM) into Next-Token Prediction (NTP) to enhance the latter’s in-context retrieval capabilities. Specifically, MEAP first randomly masks a small fraction of input tokens and then directly performs the standard next-token prediction autoregressive using a decoder-only Transformer. MEAP eliminates the need for bidirectional attention or encoder-decoder architectures for MLM, incurring no additional computational overhead during pre-training or inference. Intensive experiments demonstrate that MEAP substantially outperforms NTP on key information retrieval and long-context reasoning tasks, while performing on par or better on commonsense reasoning tasks. The benefits of MEAP also extend to supervised fine-tuning, where it shows remarkable advantages in lost-in-the-middle scenarios, outperforming NTP by 11.77 percentage points. Our analysis indicates that MEAP’s effectiveness arises from its ability to promote more distinguishable attention scores by concentrating on a reduced set of non-masked tokens. This mechanism improves the model’s focus on task-relevant signals while mitigating the influence of peripheral context. These findings position MEAP as a promising training paradigm for large language models.

Method: The study tests whether statistical properties of parse trees align with assumptions in the power-law argument. Using English and Japanese corpora, researchers analyze mutual information, deviations from probabilistic context-free grammars (PCFGs), and other properties in natural language parse trees and in PCFGs that approximate these parse trees.

Result: Results indicate that the assumptions do not hold for syntactic structures. The proposed argument is difficult to apply not only to sentences by human adults but also to other domains, suggesting a need to reconsider the relationship between power laws and hierarchical structures.

Conclusion: The study challenges the interpretation that power-law correlations in language necessarily indicate hierarchical structures, highlighting the need for more careful empirical validation of statistical arguments about language structure.

Abstract: Statistical analysis of corpora provides an approach to quantitatively investigate natural languages. This approach has revealed that several power laws consistently emerge across different corpora and languages, suggesting universal mechanisms underlying languages. In particular, the power-law decay of correlations has been interpreted as evidence of underlying hierarchical structures in syntax, semantics, and discourse. This perspective has also been extended beyond corpora produced by human adults, including child speech, birdsong, and chimpanzee action sequences. However, the argument supporting this interpretation has not been empirically tested in natural languages. To address this gap, the present study examines the validity of the argument for syntactic structures. Specifically, we test whether the statistical properties of parse trees align with the assumptions in the argument. Using English and Japanese corpora, we analyze the mutual information, deviations from probabilistic context-free grammars (PCFGs), and other properties in natural language parse trees, as well as in the PCFG that approximates these parse trees. Our results indicate that the assumptions do not hold for syntactic structures and that it is difficult to apply the proposed argument not only to sentences by human adults but also to other domains, highlighting the need to reconsider the relationship between the power law and hierarchical structures.

Method: Created Re^2 dataset from 24 conferences and 21 workshops on OpenReview, containing 19,926 initial submissions, 70,668 review comments, and 53,818 rebuttals. Framed rebuttal/discussion as multi-turn conversation to support both static review tasks and dynamic interactive LLM assistants.

Result: Largest consistency-ensured peer review and rebuttal dataset available, with initial submissions (not revised versions) and multi-turn conversation structure for rebuttal interactions.

Conclusion: Re^2 dataset addresses limitations of existing peer review data and enables better LLM-based tools to help authors self-evaluate before submission and reduce review burden.

Abstract: Peer review is a critical component of scientific progress in the fields like AI, but the rapid increase in submission volume has strained the reviewing system, which inevitably leads to reviewer shortages and declines review quality. Besides the growing research popularity, another key factor in this overload is the repeated resubmission of substandard manuscripts, largely due to the lack of effective tools for authors to self-evaluate their work before submission. Large Language Models (LLMs) show great promise in assisting both authors and reviewers, and their performance is fundamentally limited by the quality of the peer review data. However, existing peer review datasets face three major limitations: (1) limited data diversity, (2) inconsistent and low-quality data due to the use of revised rather than initial submissions, and (3) insufficient support for tasks involving rebuttal and reviewer-author interactions. To address these challenges, we introduce the largest consistency-ensured peer review and rebuttal dataset named Re^2, which comprises 19,926 initial submissions, 70,668 review comments, and 53,818 rebuttals from 24 conferences and 21 workshops on OpenReview. Moreover, the rebuttal and discussion stage is framed as a multi-turn conversation paradigm to support both traditional static review tasks and dynamic interactive LLM assistants, providing more practical guidance for authors to refine their manuscripts and helping alleviate the growing review burden. Our data and code are available in https://anonymous.4open.science/r/ReviewBench_anon/.

Method: Two-stage adaptive algorithm: 1) exploratory phase estimates reward distribution for each prompt using small budget, 2) adaptive allocation stage uses these estimates to distribute remaining compute budget efficiently across prompts based on their alignment difficulty.

Result: The adaptive strategy outperforms uniform allocation with same inference budget across 12 LM/RM pairs and 50 different prompt batches from AlpacaEval, HH-RLHF, and PKU-SafeRLHF datasets. Remains competitive against uniform allocations with 20% larger budgets and improves with larger batch sizes.

Conclusion: Prompt-adaptive Best-of-N alignment provides a simple, practical approach for more efficient test-time alignment that can reduce computational costs while maintaining or improving alignment performance.

Abstract: Recent advances in test-time alignment methods, such as Best-of-N sampling, offer a simple and effective way to steer language models (LMs) toward preferred behaviors using reward models (RM). However, these approaches can be computationally expensive, especially when applied uniformly across prompts without accounting for differences in alignment difficulty. In this work, we propose a prompt-adaptive strategy for Best-of-N alignment that allocates inference-time compute more efficiently. Motivated by latency concerns, we develop a two-stage algorithm: an initial exploratory phase estimates the reward distribution for each prompt using a small exploration budget, and a second stage adaptively allocates the remaining budget using these estimates. Our method is simple, practical, and compatible with any LM-RM combination. Empirical results on prompts from the AlpacaEval, HH-RLHF, and PKU-SafeRLHF datasets for 12 LM/RM pairs and 50 different batches of prompts show that our adaptive strategy outperforms the uniform allocation with the same inference budget. Moreover, we show that our adaptive strategy remains competitive against uniform allocations with 20 percent larger inference budgets and improves in performance as the batch size grows.

Method: Proposes Token Distillation, which distills representations obtained using the original tokenization to quickly learn high-quality input embeddings for new tokens without requiring extensive retraining.

Result: Experimental results across various open-weight models show Token Distillation outperforms even strong baselines in learning embeddings for new tokens.

Conclusion: Token Distillation provides an efficient method for extending language model vocabularies by learning high-quality embeddings for new tokens through distillation, addressing the limitations of static vocabularies for specialized domains.

Abstract: Current language models rely on static vocabularies determined at pretraining time, which can lead to decreased performance and increased computational cost for domains underrepresented in the original vocabulary. New tokens can be added to solve this problem, when coupled with a good initialization for their new embeddings. However, existing embedding initialization methods require expensive further training or pretraining of additional modules. In this paper, we propose Token Distillation and show that by distilling representations obtained using the original tokenization, we can quickly learn high-quality input embeddings for new tokens. Experimental results with a wide range of open-weight models show that Token Distillation outperforms even strong baselines.

Method: Created RealWorldQuestioning benchmark dataset from real user questions across four domains; defined and studied “entropy bias” (information discrepancy); tested four LLMs; evaluated responses qualitatively/quantitatively using ChatGPT-4o as “LLM-as-judge”; proposed iterative prompt-based debiasing approach merging gendered responses.

Result: No significant bias at category level, but substantial differences at individual question level (biases cancel out across categories); debiasing approach produced higher information content than both gendered variants in 78% of cases and balanced integration in remaining cases.

Conclusion: LLMs show concerning gender bias at granular question level despite appearing fair at category level; simple prompt-based debiasing can effectively mitigate bias and improve response quality.

Abstract: We investigate the existence and persistence of a specific type of gender bias in some of the popular LLMs and contribute a new benchmark dataset, RealWorldQuestioning (released on HuggingFace ), developed from real-world questions across four key domains in business and health contexts: education, jobs, personal financial management, and general health. We define and study entropy bias, which we define as a discrepancy in the amount of information generated by an LLM in response to real questions users have asked. We tested this using four different LLMs and evaluated the generated responses both qualitatively and quantitatively by using ChatGPT-4o (as “LLM-as-judge”). Our analyses (metric-based comparisons and “LLM-as-judge” evaluation) suggest that there is no significant bias in LLM responses for men and women at a category level. However, at a finer granularity (the individual question level), there are substantial differences in LLM responses for men and women in the majority of cases, which “cancel” each other out often due to some responses being better for males and vice versa. This is still a concern since typical users of these tools often ask a specific question (only) as opposed to several varied ones in each of these common yet important areas of life. We suggest a simple debiasing approach that iteratively merges the responses for the two genders to produce a final result. Our approach demonstrates that a simple, prompt-based debiasing strategy can effectively debias LLM outputs, thus producing responses with higher information content than both gendered variants in 78% of the cases, and consistently achieving a balanced integration in the remaining cases.

Method: The method uses: 1) target language corpora, 2) existing open-weight multilingual base and instructed backbone LLMs, and 3) synthetically generated instructions sampled from the instructed backbone. The approach systematically studies different combinations of these components, with experiments conducted for Basque language.

Result: Results show that target language corpora are essential, synthetic instructions yield robust models, and using an instruction-tuned model as backbone outperforms using a base non-instructed model. Scaling to Llama 3.1 Instruct 70B produces models competitive with much larger frontier models for Basque, without using any Basque instructions.

Conclusion: The paper demonstrates an effective approach for adapting LLMs to low-resource languages using synthetic instructions and existing multilingual models, providing a practical solution for languages lacking large instruction datasets. All resources are released for reproducibility.

Abstract: Instructing language models with user intent requires large instruction datasets, which are only available for a limited set of languages. In this paper, we explore alternatives to conventional instruction adaptation pipelines in low-resource scenarios. We assume a realistic scenario for low-resource languages, where only the following are available: corpora in the target language, existing open-weight multilingual base and instructed backbone LLMs, and synthetically generated instructions sampled from the instructed backbone. We present a comprehensive set of experiments for Basque that systematically study different combinations of these components evaluated on benchmarks and human preferences from 1,680 participants. Our conclusions show that target language corpora are essential, with synthetic instructions yielding robust models, and, most importantly, that using as backbone an instruction-tuned model outperforms using a base non-instructed model. Scaling up to Llama 3.1 Instruct 70B as backbone, our model comes near frontier models of much larger sizes for Basque, without using any Basque instructions. We release code, models, instruction datasets, and human preferences to support full reproducibility in future research on low-resource language adaptation. https://github.com/hitz-zentroa/latxa-instruct

Method: A modular architecture with four LLM-powered agents (Master, Planner, Executor, Writer) that collaborate through coordinated workflows to evaluate query complexity, decompose problems, orchestrate tool usage, execute tasks, and synthesize content.

Result: A comprehensive blueprint for AI search systems with systematic methodologies for task planning, tool integration, execution strategies, retrieval-augmented generation, and LLM inference optimizations.

Conclusion: The work provides foundational components to inform development of trustworthy, adaptive, and scalable AI search systems through the proposed AI Search Paradigm.

Abstract: In this paper, we introduce the AI Search Paradigm, a comprehensive blueprint for next-generation search systems capable of emulating human information processing and decision-making. The paradigm employs a modular architecture of four LLM-powered agents (Master, Planner, Executor and Writer) that dynamically adapt to the full spectrum of information needs, from simple factual queries to complex multi-stage reasoning tasks. These agents collaborate dynamically through coordinated workflows to evaluate query complexity, decompose problems into executable plans, and orchestrate tool usage, task execution, and content synthesis. We systematically present key methodologies for realizing this paradigm, including task planning and tool integration, execution strategies, aligned and robust retrieval-augmented generation, and efficient LLM inference, spanning both algorithmic techniques and infrastructure-level optimizations. By providing an in-depth guide to these foundational components, this work aims to inform the development of trustworthy, adaptive, and scalable AI search systems.

Method: The authors build upon Stirling’s unified framework from ecology/economics (variety, balance, disparity dimensions), survey over 300 recent diversity-related papers from ACL Anthology, and construct an NLP-specific framework with 4 perspectives.

Result: The analysis increases comparability of diversity approaches in NLP, reveals emerging trends, and allows formulation of recommendations for the field through a systematic framework.

Conclusion: The paper provides a much-needed conceptual framework for diversity in NLP, helping to standardize approaches and improve consistency across research in this important area.

Abstract: The concept of diversity has received increasing attention in natural language processing (NLP) in recent years. It became an advocated property of datasets and systems, and many measures are used to quantify it. However, it is often addressed in an ad hoc manner, with few explicit justifications of its endorsement and many cross-paper inconsistencies. There have been very few attempts to take a step back and understand the conceptualization of diversity in NLP. To address this fragmentation, we take inspiration from other scientific fields where the concept of diversity has been more thoroughly conceptualized. We build upon Stirling (2007), a unified framework adapted from ecology and economics, which distinguishes three dimensions of diversity: variety, balance, and disparity. We survey over 300 recent diversity-related papers from ACL Anthology and build an NLP-specific framework with 4 perspectives: why diversity is important, what diversity is measured on, where it is measured, and how. Our analysis increases comparability of approaches to diversity in NLP, reveals emerging trends and allows us to formulate recommendations for the field.

Method: Introduced CoRE, a large-scale benchmark based on cognitive appraisal theory to probe implicit cognitive structures LLMs use when interpreting emotionally charged situations. Assessed alignment with human appraisal patterns, internal consistency, cross-model generalization, and robustness to contextual variation.

Result: LLMs capture systematic relations between cognitive appraisals and emotions but show misalignment with human judgments and instability across contexts.

Conclusion: While LLMs demonstrate some ability to reason about emotions through cognitive dimensions, they lack alignment with human emotional reasoning and show contextual instability, highlighting limitations in current emotion understanding capabilities.

Abstract: Affective computing seeks to support the holistic development of artificial intelligence by enabling machines to engage with human emotion. Recent foundation models, particularly large language models (LLMs), have been trained and evaluated on emotion-related tasks, typically using supervised learning with discrete emotion labels. Such evaluations largely focus on surface phenomena, such as recognizing expressed or evoked emotions, leaving open whether these systems reason about emotion in cognitively meaningful ways. Here we ask whether LLMs can reason about emotions through underlying cognitive dimensions rather than labels alone. Drawing on cognitive appraisal theory, we introduce CoRE, a large-scale benchmark designed to probe the implicit cognitive structures LLMs use when interpreting emotionally charged situations. We assess alignment with human appraisal patterns, internal consistency, cross-model generalization, and robustness to contextual variation. We find that LLMs capture systematic relations between cognitive appraisals and emotions but show misalignment with human judgments and instability across contexts.

Method: Two studies: 1) English color-naming study comparing LLMs’ complexity and English-alignment; 2) Simulated cultural evolution using Iterated in-Context Language Learning (IICLL) to test if LLMs exhibit human-like inductive bias toward IB-efficiency.

Result: Larger instruction-tuned models achieve better English-alignment and IB-efficiency. Only Gemini 2.0 recapitulates the wide range of near-optimal IB-tradeoffs observed in humans, while other models converge to low-complexity solutions.

Conclusion: Human-aligned semantic categories can emerge in LLMs via the same Information Bottleneck principles that underlie semantic efficiency in humans, but only models with strong in-context capabilities achieve full human-like efficiency.

Abstract: Converging evidence suggests that human systems of semantic categories achieve near-optimal compression via the Information Bottleneck (IB) complexity-accuracy tradeoff. Large language models (LLMs) are not trained for this objective, which raises the question: are LLMs capable of evolving efficient human-aligned semantic systems? To address this question, we focus on color categorization – a key testbed of cognitive theories of categorization with uniquely rich human data – and replicate with LLMs two influential human studies. First, we conduct an English color-naming study, showing that LLMs vary widely in their complexity and English-alignment, with larger instruction-tuned models achieving better alignment and IB-efficiency. Second, to test whether these LLMs simply mimic patterns in their training data or actually exhibit a human-like inductive bias toward IB-efficiency, we simulate cultural evolution of pseudo color-naming systems in LLMs via a method we refer to as Iterated in-Context Language Learning (IICLL). We find that akin to humans, LLMs iteratively restructure initially random systems towards greater IB-efficiency. However, only a model with strongest in-context capabilities (Gemini 2.0) is able to recapitulate the wide range of near-optimal IB-tradeoffs observed in humans, while other state-of-the-art models converge to low-complexity solutions. These findings demonstrate how human-aligned semantic categories can emerge in LLMs via the same fundamental principle that underlies semantic efficiency in humans.

Method: Prove linear separability of standard subregular language classes using their deciding predicates, conduct synthetic experiments under noise-free conditions, and apply to real-data experiments on English morphology.

Result: Perfect separability confirmed in synthetic experiments; learned features align with well-known linguistic constraints in English morphology experiments; finite observability established.

Conclusion: The subregular hierarchy offers a rigorous, interpretable foundation for natural language modeling with guaranteed learnability using simple linear models.

Abstract: We prove that all standard subregular language classes are linearly separable when represented by their deciding predicates. This establishes finite observability and guarantees learnability with simple linear models. Synthetic experiments confirm perfect separability under noise-free conditions, while real-data experiments on English morphology show that learned features align with well-known linguistic constraints. These results demonstrate that the subregular hierarchy provides a rigorous and interpretable foundation for modeling natural language structure. Our code used in real-data experiments is available at https://github.com/UTokyo-HayashiLab/subregular.

Method: SPELL integrates three cyclical roles within a single model: questioner generates questions from documents with reference answers, responder solves these questions, and verifier evaluates semantic equivalence to produce reward signals. Includes automated curriculum for document length and adaptive reward function.

Result: Extensive experiments on six long-context benchmarks show consistent performance improvements across diverse LLMs, outperforming equally sized models fine-tuned on large-scale annotated data. Achieves average 7.6-point gain in pass@8 on Qwen3-30B-A3B-Thinking.

Conclusion: SPELL enables scalable, label-free optimization for long-context reasoning, showing promise for scaling to more capable models and raising performance ceilings of strong reasoning models.

Abstract: Progress in long-context reasoning for large language models (LLMs) has lagged behind other recent advances. This gap arises not only from the intrinsic difficulty of processing long texts, but also from the scarcity of reliable human annotations and programmatically verifiable reward signals. In this paper, we propose SPELL, a multi-role self-play reinforcement learning framework that enables scalable, label-free optimization for long-context reasoning. SPELL integrates three cyclical roles-questioner, responder, and verifier-within a single model to enable continual self-improvement. The questioner generates questions from raw documents paired with reference answers; the responder learns to solve these questions based on the documents; and the verifier evaluates semantic equivalence between the responder’s output and the questioner’s reference answer, producing reward signals to guide continual training. To stabilize training, we introduce an automated curriculum that gradually increases document length and a reward function that adapts question difficulty to the model’s evolving capabilities. Extensive experiments on six long-context benchmarks show that SPELL consistently improves performance across diverse LLMs and outperforms equally sized models fine-tuned on large-scale annotated data. Notably, SPELL achieves an average 7.6-point gain in pass@8 on the strong reasoning model Qwen3-30B-A3B-Thinking, raising its performance ceiling and showing promise for scaling to even more capable models. Our code is available at https://github.com/Tongyi-Zhiwen/Qwen-Doc.

Method: Co-created evaluation pipeline with civil-society organizations and community members, enabling scalable automated benchmarking through culturally-aware, community-driven approach.

Result: Demonstrated approach in Indian healthcare domain, showing how multilingual LLMs handle nuanced community health queries while providing scalable contextual evaluation.

Conclusion: Community-driven evaluation offers a pathway for more inclusive and contextually grounded LLM assessment, particularly important for domains like healthcare.

Abstract: Large Language Models (LLMs) are typically evaluated through general or domain-specific benchmarks testing capabilities that often lack grounding in the lived realities of end users. Critical domains such as healthcare require evaluations that extend beyond artificial or simulated tasks to reflect the everyday needs, cultural practices, and nuanced contexts of communities. We propose Samiksha, a community-driven evaluation pipeline co-created with civil-society organizations (CSOs) and community members. Our approach enables scalable, automated benchmarking through a culturally aware, community-driven pipeline in which community feedback informs what to evaluate, how the benchmark is built, and how outputs are scored. We demonstrate this approach in the health domain in India. Our analysis highlights how current multilingual LLMs address nuanced community health queries, while also offering a scalable pathway for contextually grounded and inclusive LLM evaluation.

Method: DataMind uses: 1) fine-grained task taxonomy with recursive easy-to-hard composition, 2) knowledge-augmented trajectory sampling with filtering, 3) dynamic training objective combining SFT and RL losses, 4) memory-frugal code-based multi-turn rollout framework.

Result: DataMind-14B achieves 71.16% average score on multiple data analysis benchmarks, outperforming proprietary baselines like DeepSeek-V3.1 and GPT-5. DataMind-7B performs best among open-source models with 68.10% score.

Conclusion: DataMind provides an effective framework for building generalist data-analytic agents, with released datasets and models (DataMind-12K, DataMind-7B/14B) to advance research in automated scientific discovery and agentic AI.

Abstract: Data-analytic agents are emerging as a key catalyst for automated scientific discovery and for the vision of Innovating AI. Current approaches, however, rely heavily on prompt engineering over proprietary models, while open-source models struggle to face diverse-format, large-scale data files and long-horizon, multi-step reasoning that real-world analytics demands. This paper introduces DataMind, a scalable data synthesis and agent training recipe designed to build generalist data-analytic agents. DataMind tackles three key challenges in building open-source data-analytic agents, including insufficient data resources, improper training strategy, and unstable code-based multi-turn rollout. Concretely, DataMind applies 1) a fine-grained task taxonomy and a recursive easy-to-hard task composition mechanism to increase the diversity and difficulty of synthesized queries; 2) a knowledge-augmented trajectory sampling strategy followed by model-based and rule-based filtering; 3) a dynamically adjustable training objective combining both SFT and RL losses; 4) a memory-frugal and stable code-based multi-turn rollout framework. Built on DataMind, we curate DataMind-12K, a high-quality trajectory set spanning diverse domains, task categories, and data file formats for data-analytic tasks. Trained on DataMind-12K, our DataMind-14B achieves state-of-the-art with an average score of 71.16% on multiple data analysis benchmarks, outperforming the strongest proprietary baselines DeepSeek-V3.1 and GPT-5. Our DataMind-7B also performs best among all open-source models with a score of 68.10%. We also incorporate some empirical insights gained from our exploratory trials into the analysis experiments, aiming to provide actionable insights about agentic training for the community. We will release DataMind-12K and DataMind-7B,14B for the community’s future research.

Method: SAFE identifies ensembling positions by considering two factors: tokenization mismatch across models and consensus in next-token probability distributions. It applies probability sharpening when ensemble distributions become overly smooth to select more confident tokens.

Result: Experiments on MATH500 and BBH benchmarks show SAFE outperforms existing methods in both accuracy and efficiency, achieving gains even when ensembling fewer than 1% of tokens.

Conclusion: Selective ensembling based on tokenization mismatch and probability consensus is crucial for effective long-form generation, and SAFE provides a stable and efficient framework for this purpose.

Abstract: Ensembling Large Language Models (LLMs) has gained attention as a promising approach to surpass the performance of individual models by leveraging their complementary strengths. In particular, aggregating models’ next-token probability distributions to select the next token has been shown to be effective in various tasks. However, while successful for short-form answers, its application to long-form generation remains underexplored. In this paper, we show that using existing ensemble methods in long-form generation requires a careful choice of ensembling positions, since the standard practice of ensembling at every token often degrades performance. We identify two key factors for determining the ensembling positions: tokenization mismatch across models and consensus in their next-token probability distributions. Based on this, we propose SAFE, (Stable And Fast LLM Ensembling), a framework that selectively ensembles by jointly considering these factors. To further improve stability, we apply a probability sharpening strategy when the ensemble distribution becomes overly smooth, enabling the selection of more confident tokens during ensembling. Our experiments on diverse benchmarks, including MATH500 and BBH, demonstrate that SAFE outperforms existing methods in both accuracy and efficiency, with gains achieved even when ensembling fewer than 1% of tokens.

Method: RECAP uses a feedback-driven loop where an initial extraction attempt is evaluated by a secondary LLM that compares output against reference passages, identifies discrepancies, and generates minimal correction hints for iterative refinement. Includes jailbreaking module to overcome alignment-induced refusals.

Result: Evaluation on EchoTrace benchmark (30+ full books) shows RECAP achieves substantial gains over single-iteration approaches. With GPT-4.1, average ROUGE-L score for copyrighted text extraction improved from 0.38 to 0.47 (24% increase).

Conclusion: RECAP effectively extracts memorized training data from LLMs through iterative refinement and jailbreaking, providing a method to audit model training content without direct data access.

Abstract: If we cannot inspect the training data of a large language model (LLM), how can we ever know what it has seen? We believe the most compelling evidence arises when the model itself freely reproduces the target content. As such, we propose RECAP, an agentic pipeline designed to elicit and verify memorized training data from LLM outputs. At the heart of RECAP is a feedback-driven loop, where an initial extraction attempt is evaluated by a secondary language model, which compares the output against a reference passage and identifies discrepancies. These are then translated into minimal correction hints, which are fed back into the target model to guide subsequent generations. In addition, to address alignment-induced refusals, RECAP includes a jailbreaking module that detects and overcomes such barriers. We evaluate RECAP on EchoTrace, a new benchmark spanning over 30 full books, and the results show that RECAP leads to substantial gains over single-iteration approaches. For instance, with GPT-4.1, the average ROUGE-L score for the copyrighted text extraction improved from 0.38 to 0.47 - a nearly 24% increase.

Method: DeCode is a training-free, model-agnostic framework that decouples content (medical facts) from delivery (contextual adaptation) to adapt existing LLMs for clinical settings without additional training.

Result: DeCode boosts zero-shot performance from 28.4% to 49.8% on OpenAI HealthBench, achieving new state-of-the-art compared to existing methods for clinical question answering.

Conclusion: The framework effectively improves LLMs’ clinical question answering by enabling contextual adaptation without retraining, making medical AI more patient-centered.

Abstract: Large language models (LLMs) exhibit strong medical knowledge and can generate factually accurate responses. However, existing models often fail to account for individual patient contexts, producing answers that are clinically correct yet poorly aligned with patients’ needs. In this work, we introduce DeCode (Decoupling Content and Delivery), a training-free, model-agnostic framework that adapts existing LLMs to produce contextualized answers in clinical settings. We evaluate DeCode on OpenAI HealthBench, a comprehensive and challenging benchmark designed to assess clinical relevance and validity of LLM responses. DeCode boosts zero-shot performance from 28.4% to 49.8% and achieves new state-of-the-art compared to existing methods. Experimental results suggest the effectiveness of DeCode in improving clinical question answering of LLMs.

Method: Factorial study investigating four factors: forecasting model choice (XGBoost, Random Forest, MLP, SARIMAX), XAI method (SHAP, LIME, no-XAI baseline), LLM selection (GPT-4o, Llama-3-8B, DeepSeek-R1), and eight prompting strategies. Evaluated 660 explanations using G-Eval with dual LLM judges and four criteria.

Result: 1) XAI provides only small improvements over no-XAI baselines, and only for expert audiences; 2) LLM choice dominates all other factors, with DeepSeek-R1 outperforming GPT-4o and Llama-3; 3) Interpretability paradox: SARIMAX yielded lower NLE quality than ML models despite higher prediction accuracy; 4) Zero-shot prompting competitive with self-consistency at 7-times lower cost; 5) Chain-of-thought hurts rather than helps.

Conclusion: LLM selection is the most critical factor for generating high-quality natural language explanations from XAI outputs, with XAI methods providing limited additional value. Practical implications suggest focusing on LLM choice and simpler prompting strategies.

Abstract: Explainable AI (XAI) methods like SHAP and LIME produce numerical feature attributions that remain inaccessible to non expert users. Prior work has shown that Large Language Models (LLMs) can transform these outputs into natural language explanations (NLEs), but it remains unclear which factors contribute to high-quality explanations. We present a systematic factorial study investigating how Forecasting model choice, XAI method, LLM selection, and prompting strategy affect NLE quality. Our design spans four models (XGBoost (XGB), Random Forest (RF), Multilayer Perceptron (MLP), and SARIMAX - comparing black-box Machine-Learning (ML) against classical time-series approaches), three XAI conditions (SHAP, LIME, and a no-XAI baseline), three LLMs (GPT-4o, Llama-3-8B, DeepSeek-R1), and eight prompting strategies. Using G-Eval, an LLM-as-a-judge evaluation method, with dual LLM judges and four evaluation criteria, we evaluate 660 explanations for time-series forecasting. Our results suggest that: (1) XAI provides only small improvements over no-XAI baselines, and only for expert audiences; (2) LLM choice dominates all other factors, with DeepSeek-R1 outperforming GPT-4o and Llama-3; (3) we observe an interpretability paradox: in our setting, SARIMAX yielded lower NLE quality than ML models despite higher prediction accuracy; (4) zero-shot prompting is competitive with self-consistency at 7-times lower cost; and (5) chain-of-thought hurts rather than helps.

Method: Comprehensive pipeline that aggregates ClaimReview feeds, scrapes full debunking articles, normalizes verdicts, enriches with structured metadata and aligned visual content, and uses LLMs and multimodal LLMs for evidence extraction and justification generation.

Result: Evaluation with G-Eval and human assessment shows the pipeline enables fine-grained comparison of fact-checking practices, facilitates development of interpretable fact-checking models, and supports multilingual, multimodal misinformation verification research.

Conclusion: The pipeline provides a foundation for robust, up-to-date, explainable, and multilingual fact-checking resources that incorporate multimodal evidence and structured annotations.

Abstract: The rapid proliferation of misinformation across online platforms underscores the urgent need for robust, up-to-date, explainable, and multilingual fact-checking resources. However, existing datasets are limited in scope, often lacking multimodal evidence, structured annotations, and detailed links between claims, evidence, and verdicts. This paper introduces a comprehensive data collection and processing pipeline that constructs multimodal fact-checking datasets in French and German languages by aggregating ClaimReview feeds, scraping full debunking articles, normalizing heterogeneous claim verdicts, and enriching them with structured metadata and aligned visual content. We used state-of-the-art large language models (LLMs) and multimodal LLMs for (i) evidence extraction under predefined evidence categories and (ii) justification generation that links evidence to verdicts. Evaluation with G-Eval and human assessment demonstrates that our pipeline enables fine-grained comparison of fact-checking practices across different organizations or media markets, facilitates the development of more interpretable and evidence-grounded fact-checking models, and lays the groundwork for future research on multilingual, multimodal misinformation verification.

Method: Collected 246K news articles from 11 media outlets across Russia, Ukraine, U.S., U.K., and China spanning February 2022 to August 2024. Articles cover three languages and include comprehensive metadata with human-evaluated annotations for stance, sentiment, and topical framing.

Result: Created DNIPRO corpus enabling systematic analysis of narrative divergence, media framing, and information warfare. Exploratory analyses revealed how media outlets construct incompatible realities through divergent attribution and topical selection without direct refutation of opposing narratives.

Conclusion: DNIPRO empowers empirical research on narrative evolution, cross-lingual information flow, and computational detection of implicit contradictions in fragmented information ecosystems, providing a valuable resource for studying geopolitical narratives and information warfare.

Abstract: We present DNIPRO, a corpus of 246K news articles from the Russo-Ukrainian war (Feb 2022 – Aug 2024) spanning eleven outlets across five nation-states (Russia, Ukraine, U.S., U.K., China) and three languages. The corpus features comprehensive metadata and human-evaluated annotations for stance, sentiment, and topical framing, enabling systematic analysis of competing geopolitical narratives. It is uniquely suited for empirical studies of narrative divergence, media framing, and information warfare. Our exploratory analyses reveal how media outlets construct incompatible realities through divergent attribution and topical selection without direct refutation of opposing narratives. DNIPRO empowers empirical research on narrative evolution, cross-lingual information flow, and computational detection of implicit contradictions in fragmented information ecosystems.

Method: Developed three attack methods: 1) logistic regression (baseline), 2) 3-layer MLP that improves reconstruction, and 3) transformer-based sequence decoder that achieves highest accuracy. Trained on 100M tokens from OpenWebText and tested on 32-token sequences.

Result: Transformer-based decoder recovers 91.2% of tokens top-1 (94.8% top-10) on 32-token sequences. 3-layer MLP achieves 63.1% top-1 accuracy. Adding noise reduces but doesn’t eliminate reconstruction. Practical leakage scenarios include distributed inference and side channels.

Conclusion: Expert selections in MoE deployments leak substantial information and should be treated as sensitive as the underlying text. The attacks demonstrate that routing decisions are highly informative for text reconstruction, raising security concerns for MoE model deployments.

Abstract: We present a text-reconstruction attack on mixture-of-experts (MoE) language models that recovers tokens from expert selections alone. In MoE models, each token is routed to a subset of expert subnetworks; we show these routing decisions leak substantially more information than previously understood. Prior work using logistic regression achieves limited reconstruction; we show that a 3-layer MLP improves this to 63.1% top-1 accuracy, and that a transformer-based sequence decoder recovers 91.2% of tokens top-1 (94.8% top-10) on 32-token sequences from OpenWebText after training on 100M tokens. These results connect MoE routing to the broader literature on embedding inversion. We outline practical leakage scenarios (e.g., distributed inference and side channels) and show that adding noise reduces but does not eliminate reconstruction. Our findings suggest that expert selections in MoE deployments should be treated as sensitive as the underlying text.

Method: Proposes Pyramid MoA, a hierarchical Mixture-of-Agents architecture with a decision-theoretic router that dynamically escalates queries based on Value of Computation theory. Establishes Probabilistic Anytime Property and extends classical monitoring framework to stochastic LLM inference.

Result: On MBPP code generation, router intercepts 81.6% of bugs. On GSM8K/MMLU, matches Oracle baseline (68.1% accuracy) with 18.4% compute savings. Zero-shot transfer to HumanEval: 81.1% accuracy with 62.7% cost savings. Preserves 58.0% Oracle ceiling on MATH 500 benchmark.

Conclusion: The framework dynamically serves as cost-cutter for low-entropy tasks and safety net for high-entropy tasks, providing an effective solution to the cost-accuracy trade-off in LLM deployment through formalized anytime computation principles.

Abstract: Large Language Models (LLMs) face a persistent trade-off between inference cost and reasoning capability. While “Oracle” models (e.g., Llama-3.3-70B) achieve state-of-the-art accuracy, they are prohibitively expensive for high-volume deployment. Smaller models (e.g., 7-9B parameters) are cost-effective but struggle with complex tasks. We observe that the emerging practice of LLM cascading and routing implicitly solves an anytime computation problem – a class of algorithms, well-studied in classical AI, that produce valid solutions immediately and improve them as additional computation is allocated. In this work, we formalize this connection and propose “Pyramid MoA”, a hierarchical Mixture-of-Agents architecture governed by a decision-theoretic router that dynamically escalates queries only when necessary. We establish a Probabilistic Anytime Property, proving that expected solution quality is monotonically non-decreasing with computational depth under identifiable conditions on router precision. We derive a generalized escalation rule from Value of Computation theory that accounts for imperfect oracles, extending the classical monitoring framework of Hansen and Zilberstein to stochastic LLM inference. On the MBPP code generation benchmark, the Consensus Router intercepts 81.6% of bugs. On the GSM8K/MMLU mathematical reasoning benchmark, the system matches the Oracle baseline of 68.1% accuracy while enabling up to 18.4% compute savings at a balanced operating point. Crucially, the router transfers zero-shot to unseen benchmarks: on HumanEval it achieves 81.1% accuracy (matching the Oracle) with 62.7% cost savings in economy mode, and on the highly complex MATH 500 benchmark it preserves the 58.0% Oracle ceiling. The framework acts dynamically: serving as an aggressive cost-cutter for low-entropy tasks and a strict safety net for high-entropy tasks.

Method: Study cross-family speculative prefill where lightweight draft models from one family (Qwen, LLaMA, DeepSeek) compress prompts for target models from different families. Use attention-based token importance estimation from prior work and evaluate diverse cross-family draft-target combinations across various tasks.

Result: Attention-based token importance estimation transfers reliably across different model families despite architectural and tokenizer differences. Cross-model compression retains 90-100% of baseline performance, sometimes slightly improving accuracy due to denoising effects, while substantially reducing time to first token.

Conclusion: Speculative prefill depends mainly on task priors and semantic structure, serving as a generalizable prompt compression primitive. This is particularly valuable for agentic systems with repeated long-context inference and heterogeneous model stacks.

Abstract: Prompt length is a major bottleneck in agentic large language model (LLM) workloads, where repeated inference steps and multi-call loops incur substantial prefill cost. Recent work on speculative prefill demonstrates that attention-based token importance estimation can enable training-free prompt compression, but this assumes the existence of a draft model that shares the same tokenizer as the target model. In practice, however, agentic pipelines frequently employ models without any smaller in-family draft model. In this work, we study cross-family speculative prefill, where a lightweight draft model from one model family is used to perform prompt compression for a target model from a different family. Using the same speculative prefill mechanism as prior work, we evaluate a range of cross-family draft-target combinations, including Qwen, LLaMA, and DeepSeek models. Across a broad diversity of tasks, we find that attention-based token importance estimation transfers reliably across different model families despite differences in model architectures and tokenizers between draft and target models. Cross-model prompt compression largely retains 90~100% of full-prompt baseline performance and, in some cases, slightly improves accuracy due to denoising effects, while delivering substantial reductions in time to first token (TTFT). These results suggest that speculative prefill depends mainly on task priors and semantic structure, thus serving as a generalizable prompt compression primitive. We discuss the implications of our findings for agentic systems, where repeated long-context inference and heterogeneous model stacks make cross-model prompt compression both necessary and practical.

Method: Created high-quality translations of challenging math problems from AceReason-Math corpus, with careful cleaning and improvement of translations across 14 languages

Result: Produced mAceReason-Math dataset with over 10,000 samples per language, covering 14 languages, to facilitate multilingual RLVR research and benchmarking

Conclusion: The dataset addresses the gap in multilingual RLVR research by providing appropriately challenging math problems for current model capabilities across multiple languages

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has been successfully applied to significantly boost the capabilities of pretrained large language models, especially in the math and logic problem domains. However, current research and available training datasets remain English-centric. While multilingual training data and benchmarks have been created in the past, they were not created with RLVR and current model capability in mind, and their level of difficulty is often too low to provide appropriate training signals for current models. To address this gap, we provide mAceReason-Math, a dataset of high-quality translations of challenging math problems sourced from a corpus specifically curated for RLVR (AceReason-Math). We further take specific care to clean and improve our translations, resulting in a coverage of 14 languages with more than 10,000 samples per language. We release the dataset to facilitate multilingual RLVR research and benchmarking in the research community.

Method: Central Supervisor dynamically decomposes user queries and delegates subtasks to modality-appropriate tools (object detection, OCR, speech transcription, etc.). Uses RouteLLM for text-only queries and SLM-assisted modality decomposition for non-text paths with adaptive routing strategies.

Result: Evaluated on 2,847 queries across 15 task categories: 72% reduction in time-to-accurate-answer, 85% reduction in conversational rework, 67% cost reduction compared to hierarchical baseline while maintaining accuracy parity.

Conclusion: Intelligent centralized orchestration fundamentally improves multimodal AI deployment economics by efficiently coordinating specialized tools across modalities.

Abstract: We present an agentic AI framework for autonomous multimodal query processing that coordinates specialized tools across text, image, audio, video, and document modalities. A central Supervisor dynamically decomposes user queries, delegates subtasks to modality-appropriate tools (e.g., object detection, OCR, speech transcription), and synthesizes results through adaptive routing strategies rather than predetermined decision trees. For text-only queries, the framework uses learned routing via RouteLLM, while non-text paths use SLM-assisted modality decomposition. Evaluated on 2,847 queries across 15 task categories, our framework achieves 72% reduction in time-to-accurate-answer, 85% reduction in conversational rework, and 67% cost reduction compared to the matched hierarchical baseline while maintaining accuracy parity. These results demonstrate that intelligent centralized orchestration fundamentally improves multimodal AI deployment economics.

Method: VQQA uses a multi-agent framework that dynamically generates visual questions and uses VLM critiques as semantic gradients to provide human-interpretable feedback, enabling closed-loop prompt optimization via natural language interface.

Result: VQQA achieves +11.57% improvement on T2V-CompBench and +8.43% on VBench2 over vanilla generation, outperforming state-of-the-art stochastic search and prompt optimization techniques.

Conclusion: VQQA provides an efficient, black-box approach to video generation optimization that replaces passive evaluation with actionable semantic feedback, significantly improving generation quality.

Abstract: Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.

Method: Proposes a decoupled kinetic paradigm using Alternating Gradient Flow (AGF) with absolute feature-space Taylor expansion to capture structural “kinetic utility”. Introduces hybrid routing framework that decouples AGF-guided offline structural search from online execution via zero-cost physical priors. Applies gradient-magnitude decoupling analysis to identify sparsity bottlenecks in Vision Transformers.

Result: AGF successfully preserves baseline functionality at extreme sparsity (75% compression on ImageNet-1K) and avoids structural collapse where traditional metrics fall below random sampling. Hybrid routing reduces heavy expert usage by ~50% (overall cost 0.92×) without sacrificing full-model accuracy on ImageNet-100.

Conclusion: The proposed kinetic paradigm effectively addresses magnitude bias in structural pruning, enabling efficient compression of vision networks while maintaining functionality. The hybrid routing framework achieves Pareto-optimal efficiency for dynamic inference in vision models.

Abstract: Efficient deep learning traditionally relies on static heuristics like weight magnitude or activation awareness (e.g., Wanda, RIA). While successful in unstructured settings, we observe a critical limitation when applying these metrics to the structural pruning of deep vision networks. These contemporary metrics suffer from a magnitude bias, failing to preserve critical functional pathways. To overcome this, we propose a decoupled kinetic paradigm inspired by Alternating Gradient Flow (AGF), utilizing an absolute feature-space Taylor expansion to accurately capture the network’s structural “kinetic utility”. First, we uncover a topological phase transition at extreme sparsity, where AGF successfully preserves baseline functionality and exhibits topological implicit regularization, avoiding the collapse seen in models trained from scratch. Second, transitioning to architectures without strict structural priors, we reveal a phenomenon of Sparsity Bottleneck in Vision Transformers (ViTs). Through a gradient-magnitude decoupling analysis, we discover that dynamic signals suffer from signal compression in converged models, rendering them suboptimal for real-time routing. Finally, driven by these empirical constraints, we design a hybrid routing framework that decouples AGF-guided offline structural search from online execution via zero-cost physical priors. We validate our paradigm on large-scale benchmarks: under a 75% compression stress test on ImageNet-1K, AGF effectively avoids the structural collapse where traditional metrics aggressively fall below random sampling. Furthermore, when systematically deployed for dynamic inference on ImageNet-100, our hybrid approach achieves Pareto-optimal efficiency. It reduces the usage of the heavy expert by approximately 50% (achieving an estimated overall cost of 0.92$\times$) without sacrificing the full-model accuracy.

Method: Introduces domain conformal bounds (DCB) to evaluate domain divergence in unknown causal factors, then proposes GenEval - a multimodal VLM approach combining foundational models (MedGemma-4B) with human knowledge via Low-Rank Adaptation (LoRA) to bridge causal gaps.

Result: Across eight DR and two SOZ datasets, GenEval achieves superior SDG performance with average accuracy of 69.2% (DR) and 81% (SOZ), outperforming strongest baselines by 9.4% and 1.8% respectively.

Conclusion: GenEval effectively bridges causal gaps for single-source domain generalization in medical imaging by combining multimodal VLMs with human knowledge through parameter-efficient adaptation.

Abstract: Generalizing image classification across domains remains challenging in critical tasks such as fundus image-based diabetic retinopathy (DR) grading and resting-state fMRI seizure onset zone (SOZ) detection. When domains differ in unknown causal factors, achieving cross-domain generalization is difficult, and there is no established methodology to objectively assess such differences without direct metadata or protocol-level information from data collectors, which is typically inaccessible. We first introduce domain conformal bounds (DCB), a theoretical framework to evaluate whether domains diverge in unknown causal factors. Building on this, we propose GenEval, a multimodal Vision Language Models (VLM) approach that combines foundational models (e.g., MedGemma-4B) with human knowledge via Low-Rank Adaptation (LoRA) to bridge causal gaps and enhance single-source domain generalization (SDG). Across eight DR and two SOZ datasets, GenEval achieves superior SDG performance, with average accuracy of 69.2% (DR) and 81% (SOZ), outperforming the strongest baselines by 9.4% and 1.8%, respectively.

Method: Proposes Referee with identity bottleneck and matching modules to model relational consistency of speaker-specific cues. Uses single one-shot example as a biometric anchor to capture fine-grained identity discrepancies between reference and test samples.

Result: Achieves state-of-the-art results on cross-dataset and cross-language evaluation protocols, including 99.4% AUC on KoDF dataset. Demonstrates strong generalization capabilities across different manipulation methods and datasets.

Conclusion: Explicitly correlating reference-based biometric priors is a key frontier for achieving generalized and reliable audiovisual forensics. The reference-aware approach effectively addresses generalization challenges in deepfake detection.

Abstract: Deepfakes generated by advanced generative models have rapidly posed serious threats, yet existing audiovisual deepfake detection approaches struggle to generalize to unseen manipulation methods. To address this, we propose a novel reference-aware audiovisual deepfake detection method, called Referee to capture fine-grained identity discrepancies. Unlike existing methods that overfit to transient spatiotemporal artifacts, Referee employs identity bottleneck and matching modules to model the relational consistency of speaker-specific cues captured by a single one-shot example as a biometric anchor. Extensive experiments on FakeAVCeleb, FaceForensics++, and KoDF demonstrate that Referee achieves state-of-the-art results on cross-dataset and cross-language evaluation protocols, including a 99.4% AUC on KoDF. These results highlight that explicitly correlating reference-based biometric priors is a key frontier for achieving generalized and reliable audiovisual forensics. The code is available at https://github.com/ewha-mmai/referee.

Method: SPARROW uses two key components: (1) Target-Specific Tracked Features (TSF) that inject temporally aligned referent cues during training, and (2) a dual-prompt design that decodes box and segmentation tokens to fuse geometric priors with semantic grounding. It operates end-to-end without external detectors via a class-agnostic SAM2-based proposer.

Result: Integrated into three open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks: up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG.

Conclusion: SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding, demonstrating effective unification of spatial accuracy and temporal stability in video MLLMs.

Abstract: Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: https://risys-lab.github.io/SPARROW

Method: Created benchmark with in-house scenario-aware captures of manually synthesized templates (26 certificate types, 5,198 images, 13,806 VQA samples). Evaluates models on 3 perception tasks, 4 credit-specific reasoning tasks, and 10 real-world robustness challenges.

Result: Gemini 3 Pro achieved best F1 score (65.16) among commercial models, Kimi-K2.5 best among open-source (60.58). Mean score 44.8±10.3 shows benchmark is non-trivial. Robustness evaluations show significant performance degradation for top models.

Conclusion: FCMBench provides a challenging, privacy-compliant benchmark for evaluating VLMs in financial credit domain, revealing current limitations and providing strong resolution for separating model capabilities.

Abstract: FCMBench is the first large-scale and privacy-compliant multimodal benchmark for real-world financial credit applications, covering tasks and robustness challenges from domain specific workflows and constraints. The current version of FCMBench covers 26 certificate types, with 5198 privacy-compliant images and 13806 paired VQA samples. It evaluates models on Perception and Reasoning tasks under real-world Robustness interferences, including 3 foundational perception tasks, 4 credit-specific reasoning tasks demanding decision-oriented visual evidence interpretation, and 10 real-world challenges for rigorous robustness stress testing. Moreover, FCMBench offers privacy-compliant realism with minimal leakage risk through in-house scenario-aware captures of manually synthesized templates, without any publicly released images. We conduct extensive evaluations of 28 state-of-the-art vision-language models spanning 14 AI companies and research institutes. Among them, Gemini 3 Pro achieves the best F1 score as a commercial model (65.16), Kimi-K2.5 achieves the best score as an open-source baseline (60.58). The mean and the std. of all tested models is 44.8 and 10.3 respectively, indicating that FCMBench is non-trivial and provides strong resolution for separating modern vision-language model capabilities. Robustness evaluations reveal that even top-performing models experience notable performance degradation under the designed challenges. We have open-sourced this benchmark to advance AI research in the credit domain and provide a domain-specific task for real-world AI applications.

Method: Uses E(3)-equivariant landmark-graph encoder with local eye geometry, binocular emphasis, auxiliary 3D gaze supervision, and closed-form ridge calibrator differentiated through episodic meta-training. Two-view canonicalization consistency loss reduces pose leakage.

Result: Achieves 5.79±1.81° RMSE after 9-point calibration vs 6.68±2.34° for Elastic Net; better on still-head queries (2.92±0.75° vs 4.45±0.30°). Lightweight: 944K params, 4.76MB ONNX, 12.58ms inference in browser.

Conclusion: EMC-Gaze provides a calibration-friendly operating point for practical deployment rather than claiming universal SOTA against heavier appearance-based systems, balancing accuracy with real-world constraints.

Abstract: Practical webcam gaze tracking is constrained not only by error, but also by calibration burden, robustness to head motion and session drift, runtime footprint, and browser use. We therefore target a deployment-oriented operating point rather than the image large-backbone regime. We cast landmark-based point-of-regard estimation as session-wise adaptation: a shared geometric encoder produces embeddings that can be aligned to a new session from a small calibration set. We present Equivariant Meta-Calibrated Gaze (EMC-Gaze), a lightweight landmark-only method combining an E(3)-equivariant landmark-graph encoder, local eye geometry, binocular emphasis, auxiliary 3D gaze-direction supervision, and a closed-form ridge calibrator differentiated through episodic meta-training. To reduce pose leakage, we use a two-view canonicalization consistency loss. The deployed predictor uses only facial landmarks and fits a per-session ridge head from brief calibration. In a fixation-style interactive evaluation over 33 sessions at 100 cm, EMC-Gaze achieves 5.79 +/- 1.81 deg RMSE after 9-point calibration versus 6.68 +/- 2.34 deg for Elastic Net; the gain is larger on still-head queries (2.92 +/- 0.75 deg vs. 4.45 +/- 0.30 deg). Across three subject holdouts of 10 subjects each, EMC-Gaze retains an advantage (5.66 +/- 0.19 deg vs. 6.49 +/- 0.33 deg). On MPIIFaceGaze with short per-session calibration, the eye-focused model reaches 8.82 +/- 1.21 deg at 16-shot calibration, ties Elastic Net at 1-shot, and outperforms it from 3-shot onward. The exported eye-focused encoder has 944,423 parameters, is 4.76 MB in ONNX, and supports calibrated browser prediction in 12.58/12.58/12.90 ms per sample (mean/median/p90) in Chromium 145 with ONNX Runtime Web. These results position EMC-Gaze as a calibration-friendly operating point rather than a universal state-of-the-art claim against heavier appearance-based systems.

Method: ABRA formulates adaptation as a geometric transport problem in the weight space of a pretrained detector. It aligns source and target domain experts to transport class-specific knowledge without requiring target domain images containing those classes.

Result: Extensive experiments across challenging domain shifts demonstrate that ABRA successfully transfers class-level specialization under multiple adverse conditions.

Conclusion: ABRA provides an effective method for domain adaptation in open-vocabulary object detection when target domains lack annotated data, enabling knowledge transfer through geometric alignment in weight space.

Abstract: Although recent Open-Vocabulary Object Detection architectures, such as Grounding DINO, demonstrate strong zero-shot capabilities, their performance degrades significantly under domain shifts. Moreover, many domains of practical interest, such as nighttime or foggy scenes, lack large annotated datasets, preventing direct fine-tuning. In this paper, we introduce Aligned Basis Relocation for Adaptation(ABRA), a method that transfers class-specific detection knowledge from a labeled source domain to a target domain where no training images containing these classes are accessible. ABRA formulates this adaptation as a geometric transport problem in the weight space of a pretrained detector, aligning source and target domain experts to transport class-specific knowledge. Extensive experiments across challenging domain shifts demonstrate that ABRA successfully teleports class-level specialization under multiple adverse conditions. Our code will be made public upon acceptance.

Method: Proposes a neuro-symbolic framework: 1) LLM dynamically extracts scene rules, 2) ASP solver performs deterministic logical arbitration for discrete driving decisions, 3) Decision-conditioned decoding transforms logical decisions into embedding vectors, 4) Differentiable Kinematic Bicycle Model generates physical baseline trajectories, 5) Neural residual corrections refine trajectories while maintaining kinematic feasibility.

Result: On nuScenes benchmark, outperforms state-of-the-art MomAD: reduces L2 mean error to 0.57m, decreases collision rate to 0.075%, and optimizes trajectory prediction consistency to 0.47m.

Conclusion: The neuro-symbolic approach successfully integrates deductive reasoning into end-to-end autonomous driving, providing interpretability, safety guarantees, and improved performance while maintaining kinematic feasibility through differentiable physical models.

Abstract: Existing end-to-end autonomous driving models rely heavily on purely data-driven inductive reasoning. This “black-box” nature leads to a lack of interpretability and absolute safety guarantees in complex, long-tail scenarios. To overcome this bottleneck, we propose a novel neuro-symbolic trajectory planning framework that seamlessly integrates rigorous deductive reasoning into end-to-end neural networks. Specifically, our framework utilizes a Large Language Model (LLM) to dynamically extract scene rules and employs an Answer Set Programming (ASP) solver for deterministic logical arbitration, generating safe and traceable discrete driving decisions. To bridge the gap between discrete symbols and continuous trajectories, we introduce a decision-conditioned decoding mechanism that transforms high-level logical decisions into learnable embedding vectors, simultaneously constraining the planning query and the physical initial velocity of a differentiable Kinematic Bicycle Model (KBM). By combining KBM-generated physical baseline trajectories with neural residual corrections, our approach inherently guarantees kinematic feasibility while ensuring a high degree of transparency. On the nuScenes benchmark, our method comprehensively outperforms the state-of-the-art baseline MomAD, reducing the L2 mean error to 0.57 m, decreasing the collision rate to 0.075%, and optimizing trajectory prediction consistency (TPC) to 0.47 m.

Method: Three-level reasoning hierarchy decomposing surgical interpretation, creation of largest surgical chain-of-thought dataset (320,000 reasoning pairs), and four-stage training pipeline progressing from supervised fine-tuning to group relative policy optimization and iterative self-improvement.

Result: Achieves highest Arena Score (64.9%) on public benchmarks vs Gemini 3.0 Pro (46.1%) and GPT-5.1 (37.9%), outperforms both proprietary reasoning models and specialized surgical VLMs on majority of tasks, with 15.2 percentage point improvement over strongest surgical baseline on external validation.

Conclusion: Surg-R1 demonstrates that hierarchical reasoning with domain-specific surgical knowledge enables both accurate predictions and interpretable reasoning chains that align with clinical expertise, advancing surgical scene understanding.

Abstract: Surgical scene understanding demands not only accurate predictions but also interpretable reasoning that surgeons can verify against clinical expertise. However, existing surgical vision-language models generate predictions without reasoning chains, and general-purpose reasoning models fail on compositional surgical tasks without domain-specific knowledge. We present Surg-R1, a surgical Vision-Language Model that addresses this gap through hierarchical reasoning trained via a four-stage pipeline. Our approach introduces three key contributions: (1) a three-level reasoning hierarchy decomposing surgical interpretation into perceptual grounding, relational understanding, and contextual reasoning; (2) the largest surgical chain-of-thought dataset with 320,000 reasoning pairs; and (3) a four-stage training pipeline progressing from supervised fine-tuning to group relative policy optimization and iterative self-improvement. Evaluation on SurgBench, comprising six public benchmarks and six multi-center external validation datasets from five institutions, demonstrates that Surg-R1 achieves the highest Arena Score (64.9%) on public benchmarks versus Gemini 3.0 Pro (46.1%) and GPT-5.1 (37.9%), outperforming both proprietary reasoning models and specialized surgical VLMs on the majority of tasks spanning instrument localization, triplet recognition, phase recognition, action recognition, and critical view of safety assessment, with a 15.2 percentage point improvement over the strongest surgical baseline on external validation.

Method: Systematic protocol spanning stitch points, stitch layer families, training losses, and downstream tasks. Introduces feature-matching loss at target model’s penultimate layer and proposes VFM Stitch Tree (VST) for sharing early layers across VFMs.

Result: Heterogeneous VFMs become reliably stitchable with proper training, especially with feature-matching loss at penultimate layer. Deep stitch points can surpass constituent models with small inference overhead. VST enables controllable accuracy-latency trade-off.

Conclusion: Model stitching can be elevated from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing representation alignment/divergence, particularly useful for multimodal LLMs leveraging multiple VFMs.

Abstract: Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model’s penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.

Method: Proposes a self-prompted instance segmentation framework leveraging Segment Anything Model (SAM) with a self-prompt generation module that automatically produces task-specific prompts. Uses Tiny Vision Transformer backbone to reduce computational cost for edge deployment.

Result: Achieves 74.6 AP^box and 74.2 AP^mask on domain-specific dataset, outperforming state-of-the-art methods in accuracy and robustness while maintaining low computational overhead.

Conclusion: Offers a deployable and efficient vision solution for automated freight train inspection, demonstrating foundation model adaptation potential in industrial-scale fault diagnosis scenarios.

Abstract: Accurate visual fault detection in freight trains remains a critical challenge for intelligent transportation system maintenance, due to complex operational environments, structurally repetitive components, and frequent occlusions or contaminations in safety-critical regions. Conventional instance segmentation methods based on convolutional neural networks and Transformers often suffer from poor generalization and limited boundary accuracy under such conditions. To address these challenges, we propose a lightweight self-prompted instance segmentation framework tailored for freight train fault detection. Our method leverages the Segment Anything Model by introducing a self-prompt generation module that automatically produces task-specific prompts, enabling effective knowledge transfer from foundation models to domain-specific inspection tasks. In addition, we adopt a Tiny Vision Transformer backbone to reduce computational cost, making the framework suitable for real-time deployment on edge devices in railway monitoring systems. We construct a domain-specific dataset collected from real-world freight inspection stations and conduct extensive evaluations. Experimental results show that our method achieves 74.6 $AP^{\text{box}}$ and 74.2 $AP^{\text{mask}}$ on the dataset, outperforming existing state-of-the-art methods in both accuracy and robustness while maintaining low computational overhead. This work offers a deployable and efficient vision solution for automated freight train inspection, demonstrating the potential of foundation model adaptation in industrial-scale fault diagnosis scenarios. Project page: https://github.com/MVME-HBUT/SAM_FTI-FDet.git

Method: Formulates SFDA as iterative bias correction: periodically identifies target images from over-predicted classes, selectively reduces predictive confidence for uncertain (high entropy) images while preserving confident predictions. Uses jointly optimized pixel-level classifier to restore discriminative localization features.

Result: Extensive experiments on cross-organ and cross-center histopathology benchmarks (glas, CAMELYON-16, CAMELYON-17) show SFDA-DeP consistently improves classification and localization over state-of-the-art SFDA baselines.

Conclusion: SFDA-DeP effectively addresses prediction bias amplification in WSOL models under domain shift through iterative bias correction, improving both classification and localization performance in histopathology applications.

Abstract: Weakly Supervised Object Localization (WSOL) models enable joint classification and region-of-interest localization in histology images using only image-class supervision. When deployed in a target domain, distributions shift remains a major cause of performance degradation, especially when applied on new organs or institutions with different staining protocols and scanner characteristics. Under stronger cross-domain shifts, WSOL predictions can become biased toward dominant classes, producing highly skewed pseudo-label distributions in the target domain. Source-Free (Unsupervised) Domain Adaptation (SFDA) methods are commonly employed to address domain shift. However, because they rely on self-training, the initial bias is reinforced over training iterations, degrading both classification and localization tasks. We identify this amplification of prediction bias as a primary obstacle to the SFDA of WSOL models in histopathology. This paper introduces \sfdadep, a method inspired by machine unlearning that formulates SFDA as an iterative process of identifying and correcting prediction bias. It periodically identifies target images from over-predicted classes and selectively reduces the predictive confidence for uncertain (high entropy) images, while preserving confident predictions. This process reduces the drift of decision boundaries and bias toward dominant classes. A jointly optimized pixel-level classifier further restores discriminative localization features under distribution shift. Extensive experiments on cross-organ and -center histopathology benchmarks (glas, CAMELYON-16, CAMELYON-17) with several WSOL models show that SFDA-DeP consistently improves classification and localization over state-of-the-art SFDA baselines. {\small Code: \href{https://anonymous.4open.science/r/SFDA-DeP-1797/}{anonymous.4open.science/r/SFDA-DeP-1797/}}

Method: Marker-based photogrammetry approach that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality 3D models of aggregate particles.

Result: The method successfully reconstructs 3D aggregate models with validated accuracy against ground-truth. Comparative analyses show significant differences between 2D and 3D morphological properties, demonstrating the importance of 3D characterization.

Conclusion: The presented approach provides a cost-effective way to obtain 3D shape information of aggregates, enabling convenient inspection, data collection, and 3D morphological analysis for construction quality control.

Abstract: Aggregates, serving as the main skeleton in assemblies of construction materials, are important functional components in various building and transportation infrastructures. They can be used in unbound layer applications, e.g. pavement base and railroad ballast, bound applications of cement concrete and asphalt concrete, and as riprap and large-sized primary crushed rocks. Information on the size and shape or morphology of aggregates can greatly facilitate the Quality Assurance/Quality Control (QA/QC) process by providing insights of aggregate behavior during composition and packing. A full 3D characterization of aggregate particle morphology is difficult both during production in a quarry and at a construction site. Many aggregate imaging approaches have been developed to quantify the particle morphology by computer vision, including 2D image-based approaches that analyze particle silhouettes and 3D scanning-based methods that require expensive devices such as 3D laser scanners or X-Ray Computed Tomography (CT) equipment. This paper presents a flexible and cost-effective photogrammetry-based approach for the 3D reconstruction of aggregate particles. The proposed approach follows a marker-based design that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality aggregate models. The accuracy of the reconstruction results was validated against ground-truth for selected aggregate samples. Comparative analyses were conducted on 2D and 3D morphological properties of the selected samples. Significant differences were found between the 2D and 3D statistics. Based on the presented approach, 3D shape information of aggregates can be obtained easily and at a low cost, thus allowing convenient aggregate inspection, data collection, and 3D morphological analysis.

Method: AbSteering introduces: (1) an abnormality-centric Chain-of-Thought scheme that enforces abnormality reasoning, and (2) a Direct Preference Optimization objective that utilizes clinically confusable abnormalities as hard negatives to enhance fine-grained discrimination.

Result: General-purpose VideoLMs possess strong transferability to high-volume medical imaging when guided by AbSteering. The framework outperforms state-of-the-art domain-specific CT foundation models pretrained with large-scale CTs, achieving superior detection sensitivity while mitigating hallucinations.

Conclusion: AbSteering demonstrates that VideoLMs can be effectively adapted for precise HRCT report generation through abnormality-centric steering, outperforming specialized medical imaging models despite being general-purpose models.

Abstract: Generating precise diagnostic reports from High-Resolution Computed Tomography (HRCT) is critical for clinical workflow, yet it remains a formidable challenge due to the high pathological diversity and spatial sparsity within 3D volumes. While Video Language Models (VideoLMs) have demonstrated remarkable spatio-temporal reasoning in general domains, their adaptability to domain-specific, high-volume medical interpretation remains underexplored. In this work, we present AbSteering, an abnormality-centric framework that steers VideoLMs toward precise HRCT report generation. Specifically, AbSteering introduces: (i) an abnormality-centric Chain-of-Thought scheme that enforces abnormality reasoning, and (ii) a Direct Preference Optimization objective that utilizes clinically confusable abnormalities as hard negatives to enhance fine-grained discrimination. Our results demonstrate that general-purpose VideoLMs possess strong transferability to high-volume medical imaging when guided by this paradigm. Notably, AbSteering outperforms state-of-the-art domain-specific CT foundation models, which are pretrained with large-scale CTs, achieving superior detection sensitivity while simultaneously mitigating hallucinations. Our data and model weights are released at https://anonymous.4open.science/r/hrct-report-generation-video-vlm-728C/

Method: SPADE-UNet conditioned on dense spatial tokens from frozen pathology foundation model (UNI), with misalignment-aware loss suite and learned stain embeddings for multiple IHC markers.

Result: Achieves SOTA distributional metrics on all four stains (HER2, Ki67, ER, PR) from single unified model on MIST dataset, and best distributional metrics on BCI dataset.

Conclusion: UNIStainNet provides tissue-level semantic guidance for stain translation, enabling single model for multiple IHC markers with systematic errors concentrated in non-tumor tissue.

Abstract: Virtual immunohistochemistry (IHC) staining from hematoxylin and eosin (H&E) images can accelerate diagnostics by providing preliminary molecular insight directly from routine sections, reducing the need for repeat sectioning when tissue is limited. Existing methods improve realism through contrastive objectives, prototype matching, or domain alignment, yet the generator itself receives no direct guidance from pathology foundation models. We present UNIStainNet, a SPADE-UNet conditioned on dense spatial tokens from a frozen pathology foundation model (UNI), providing tissue-level semantic guidance for stain translation. A misalignment-aware loss suite preserves stain quantification accuracy, and learned stain embeddings enable a single model to serve multiple IHC markers simultaneously. On MIST, UNIStainNet achieves state-of-the-art distributional metrics on all four stains (HER2, Ki67, ER, PR) from a single unified model, where prior methods typically train separate per-stain models. On BCI, it also achieves the best distributional metrics. A tissue-type stratified failure analysis reveals that remaining errors are systematic, concentrating in non-tumor tissue. Code is available at https://github.com/facevoid/UNIStainNet.

Method: GDO computes six sample descriptors for each candidate and constructs optimized training subsets for different goals, enabling efficient one-epoch training with Qwen3-VL-8B-Instruct model.

Result: GDO achieves faster convergence and higher accuracy than baseline with far fewer samples: 35.4k vs 512k on MVBench (+1.38%), 26.6k vs 512k on VideoMME (+1.67%), 27.3k vs 512k on MLVU (+3.08%), and 34.7k vs 512k on LVBench (+0.84%).

Conclusion: GDO provides an effective goal-driven data optimization framework for multimodal instruction tuning that enables faster convergence with fewer training samples under fixed protocols.

Abstract: Multimodal instruction tuning is often compute-inefficient because training budgets are spread across large mixed image-video pools whose utility is highly uneven. We present Goal-Driven Data Optimization (GDO), a framework that computes six sample descriptors for each candidate and constructs optimized 1$\times$ training subsets for different goals. Under a fixed one-epoch Qwen3-VL-8B-Instruct training and evaluation recipe on 8 H20 GPUs, GDO uses far fewer training samples than the Uni-10x baseline while converging faster and achieving higher accuracy. Relative to the fixed 512k-sample Uni-10x baseline, GDO reaches the Uni-10x reference after 35.4k samples on MVBench, 26.6k on VideoMME, 27.3k on MLVU, and 34.7k on LVBench, while improving Accuracy by +1.38, +1.67, +3.08, and +0.84 percentage points, respectively. The gains are largest on MVBench and MLVU, while LVBench improves more modestly, consistent with its ultra-long-video setting and the mismatch between that benchmark and the short-video/image-dominant training pool. Across MinLoss, Diverse, Temp, and Temp+, stronger temporal emphasis yields steadily better long-video understanding behavior. Overall, GDO provides a goal-driven data optimization framework that enables faster convergence with fewer training samples under a fixed training protocol. Code is available at https://github.com/rujiewu/GDO.

Method: Uses vision-language models (VLMs) to generate textual descriptions of key objects from degraded images, then remaps these descriptions back to produce spatial semantic guidance maps. These maps steer UIE networks through dual-guidance combining cross-attention and explicit alignment loss.

Result: When applied to different UIE baselines, the method significantly boosts performance on perceptual quality metrics and enhances performance on detection and segmentation tasks.

Conclusion: The proposed semantic-guided approach effectively improves underwater image enhancement by focusing restoration on semantic-sensitive regions, ensuring faithful restoration of key object features and better adaptation to downstream vision tasks.

Abstract: In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.

Method: Proposes a coarse-to-fine pipeline (Text → Layout → Image) inspired by human “planning before writing” cognitive process. Uses a single Multimodal Diffusion Transformer with two stages: 1) spatial planning stage predicts character bounding boxes for global arrangement, 2) content synthesis stage uses flow-matching to render high-fidelity brushwork using the layout as geometric prompt.

Result: Achieves state-of-the-art generation quality while enabling versatile downstream capabilities. Supports controllable semantic re-planning where users can resize/reposition characters with automatic harmonization of surrounding space and brush momentum. Framework extensible to artifact restoration and forensic analysis.

Conclusion: CalliMaster provides a comprehensive tool for digital cultural heritage by resolving the conflict between spatial planning and content synthesis through disentangled representation, enabling both high-quality generation and flexible editing capabilities.

Abstract: Page-level calligraphy synthesis requires balancing glyph precision with layout composition. Existing character models lack spatial context, while page-level methods often compromise brushwork detail. In this paper, we present \textbf{CalliMaster}, a unified framework for controllable generation and editing that resolves this conflict by decoupling spatial planning from content synthesis. Inspired by the human cognitive process of ``planning before writing’’, we introduce a coarse-to-fine pipeline \textbf{(Text $\rightarrow$ Layout $\rightarrow$ Image)} to tackle the combinatorial complexity of page-scale synthesis. Operating within a single Multimodal Diffusion Transformer, a spatial planning stage first predicts character bounding boxes to establish the global spatial arrangement. This intermediate layout then serves as a geometric prompt for the content synthesis stage, where the same network utilizes flow-matching to render high-fidelity brushwork. Beyond achieving state-of-the-art generation quality, this disentanglement supports versatile downstream capabilities. By treating the layout as a modifiable constraint, CalliMaster enables controllable semantic re-planning: users can resize or reposition characters while the model automatically harmonizes the surrounding void space and brush momentum. Furthermore, we demonstrate the framework’s extensibility to artifact restoration and forensic analysis, providing a comprehensive tool for digital cultural heritage.

Method: Proposes Surgical Action World (SAW) using conditional video diffusion with four lightweight conditioning signals: language prompts for tool-action context, reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. Reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis with depth consistency loss for geometric plausibility without requiring depth at inference.

Result: Achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Demonstrates downstream utility: (1) improves surgical action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) through data augmentation, (2) enables visually faithful surgical simulation from simulator-derived trajectories.

Conclusion: SAW represents a significant step toward surgical action world modeling, enabling realistic surgical video generation with lightweight conditioning signals that addresses key challenges in surgical AI and simulation while maintaining temporal consistency and visual quality.

Abstract: A surgical world model capable of generating realistic surgical action videos with precise control over tool-tissue interactions can address fundamental challenges in surgical AI and simulation – from data scarcity and rare event synthesis to bridging the sim-to-real gap for surgical automation. However, current video generation methods, the very core of such surgical world models, require expensive annotations or complex structured intermediates as conditioning signals at inference, limiting their scalability. Other approaches exhibit limited temporal consistency across complex laparoscopic scenes and do not possess sufficient realism. We propose Surgical Action World (SAW) – a step toward surgical action world modeling through video diffusion conditioned on four lightweight signals: language prompts encoding tool-action context, a reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. We design a conditional video diffusion approach that reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis. The backbone diffusion model is fine-tuned on a custom-curated dataset of 12,044 laparoscopic clips with lightweight spatiotemporal conditioning signals, leveraging a depth consistency loss to enforce geometric plausibility without requiring depth at inference. SAW achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Furthermore, we demonstrate its downstream utility for (a) surgical AI, where augmenting rare actions with SAW-generated videos improves action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) on real test data, and (b) surgical simulation, where rendering tool-tissue interaction videos from simulator-derived trajectory points toward a visually faithful simulation engine.

Method: Models device-specific degradations through calibration, unprocesses publicly available rendered images into the RAW domain of different smartphones, and trains a single-image RAW-to-RGB SR model using these image pairs.

Result: Accurate degradation modeling leads to noticeable improvements, with the SR model outperforming baselines trained on large pools of arbitrarily chosen degradations when evaluated on real data from held-out devices.

Conclusion: Principled and carefully designed degradation modeling enhances SR performance in real-world conditions, demonstrating the importance of device-specific calibration over generic degradation priors.

Abstract: Digital zoom on smartphones relies on learning-based super-resolution (SR) models that operate on RAW sensor images, but obtaining sensor-specific training data is challenging due to the lack of ground-truth images. Synthetic data generation via ``unprocessing’’ pipelines offers a potential solution by simulating the degradations that transform high-resolution (HR) images into their low-resolution (LR) counterparts. However, these pipelines can introduce domain gaps due to incomplete or unrealistic degradation modeling. In this paper, we demonstrate that principled and carefully designed degradation modeling can enhance SR performance in real-world conditions. Instead of relying on generic priors for camera blur and noise, we model device-specific degradations through calibration and unprocess publicly available rendered images into the RAW domain of different smartphones. Using these image pairs, we train a single-image RAW-to-RGB SR model and evaluate it on real data from a held-out device. Our experiments show that accurate degradation modeling leads to noticeable improvements, with our SR model outperforming baselines trained on large pools of arbitrarily chosen degradations.

Method: Proposes InterEdit, a synchronized classifier-free conditional diffusion model with two key components: 1) Semantic-Aware Plan Token Alignment using learnable tokens to capture high-level interaction cues, and 2) Interaction-Aware Frequency Token Alignment using DCT and energy pooling to model periodic motion dynamics.

Result: InterEdit improves text-to-motion consistency and edit fidelity, achieving state-of-the-art performance on the proposed Text-guided Multi-human Motion Editing (TMME) benchmark. The method outperforms existing approaches in multi-person motion editing tasks.

Conclusion: The paper successfully addresses the challenging task of multi-person 3D motion editing by introducing a new dataset, benchmark, and diffusion-based framework that effectively captures interaction dynamics through semantic and frequency alignment strategies.

Abstract: Text-guided 3D motion editing has seen success in single-person scenarios, but its extension to multi-person settings is less explored due to limited paired data and the complexity of inter-person interactions. We introduce the task of multi-person 3D motion editing, where a target motion is generated from a source and a text instruction. To support this, we propose InterEdit3D, a new dataset with manual two-person motion change annotations, and a Text-guided Multi-human Motion Editing (TMME) benchmark. We present InterEdit, a synchronized classifier-free conditional diffusion model for TMME. It introduces Semantic-Aware Plan Token Alignment with learnable tokens to capture high-level interaction cues and an Interaction-Aware Frequency Token Alignment strategy using DCT and energy pooling to model periodic motion dynamics. Experiments show that InterEdit improves text-to-motion consistency and edit fidelity, achieving state-of-the-art TMME performance. The dataset and code will be released at https://github.com/YNG916/InterEdit.

Method: Proposes Naïve PAINE that predicts numerical image quality from initial noise and prompt using T2I preference benchmarks, selects high-quality noises, and feeds them to diffusion models for generation.

Result: Experimental results show Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks for improving generative quality.

Conclusion: Naïve PAINE provides lightweight quality feedback that seamlessly integrates into existing diffusion model pipelines to reduce the “gambler’s burden” of multiple generation attempts.

Abstract: Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler’s burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Naïve PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Naïve PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Naïve PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks.

Method: Two co-designed components: 1) Memory Tokens that continuously compress past keys into dual long-term and short-term streams via exponential moving averages, maintaining global identity and recent dynamics within fixed-size cache; 2) Online RoPE Indexing that caches unrotated keys and applies positional embeddings dynamically at attention time, ensuring aggregation free of conflicting positional phases.

Result: Extensive experiments show MemRoPE outperforms existing methods in temporal coherence, visual fidelity, and subject consistency across minute- to hour-scale video generation.

Conclusion: MemRoPE enables high-quality long-horizon video generation with consistent identity and motion through training-free memory mechanisms and dynamic positional encoding.

Abstract: Autoregressive diffusion enables real-time frame streaming, yet existing sliding-window caches discard past context, causing fidelity degradation, identity drift, and motion stagnation over long horizons. Current approaches preserve a fixed set of early tokens as attention sinks, but this static anchor cannot reflect the evolving content of a growing video. We introduce MemRoPE, a training-free framework with two co-designed components. Memory Tokens continuously compress all past keys into dual long-term and short-term streams via exponential moving averages, maintaining both global identity and recent dynamics within a fixed-size cache. Online RoPE Indexing caches unrotated keys and applies positional embeddings dynamically at attention time, ensuring the aggregation is free of conflicting positional phases. These two mechanisms are mutually enabling: positional decoupling makes temporal aggregation well-defined, while aggregation makes fixed-size caching viable for unbounded generation. Extensive experiments validate that MemRoPE outperforms existing methods in temporal coherence, visual fidelity, and subject consistency across minute- to hour-scale generation.

Method: Uses patch-based Masked Image Modeling (MIM) to pre-train a 3D U-Net encoder on 1,206 unlabeled CT volumes, then applies semi-supervised learning with consistency regularization for 3D injury detection using VDETR with Vertex Relative Position Encoding, and multi-label injury classification with frozen encoder features.

Result: Achieved 56.57% validation mAP@0.50 and 45.30% test mAP@0.50 for detection with only 144 labeled samples (115% improvement over supervised-only), and 94.07% test accuracy for classification across seven injury categories with 2,244 labeled samples using frozen encoder.

Conclusion: Self-supervised pre-training combined with semi-supervised learning effectively addresses label scarcity in medical imaging, enabling robust 3D object detection with limited annotations through transferable self-supervised features.

Abstract: Accurate detection and localization of traumatic injuries in abdominal CT scans remains a critical challenge in emergency radiology, primarily due to severe scarcity of annotated medical data. This paper presents a label-efficient approach combining self-supervised pre-training with semi-supervised detection for 3D medical image analysis. We employ patch-based Masked Image Modeling (MIM) to pre-train a 3D U-Net encoder on 1,206 CT volumes without annotations, learning robust anatomical representations. The pretrained encoder enables two downstream clinical tasks: 3D injury detection using VDETR with Vertex Relative Position Encoding, and multi-label injury classification. For detection, semi-supervised learning with 2,000 unlabeled volumes and consistency regularization achieves 56.57% validation mAP@0.50 and 45.30% test mAP@0.50 with only 144 labeled training samples, representing a 115% improvement over supervised-only training. For classification, expanding to 2,244 labeled samples yields 94.07% test accuracy across seven injury categories using only a frozen encoder, demonstrating immediately transferable self-supervised features. Our results validate that self-supervised pre-training combined with semi-supervised learning effectively addresses label scarcity in medical imaging, enabling robust 3D object detection with limited annotations.

Method: PISE combines adjoint operator initialization with semantic guidance in a physics-informed deep ghost imaging framework to enhance edge perception at low sampling rates.

Result: PISE improves classification accuracy by 2.57% and reduces variance by 9x at 5% sampling rate compared to baseline methods.

Conclusion: The proposed framework effectively addresses low-bandwidth edge perception challenges by integrating physics-informed approaches with deep learning for improved performance in resource-constrained scenarios.

Abstract: We propose PISE, a physics-informed deep ghost imaging framework for low-bandwidth edge perception. By combining adjoint operator initialization with semantic guidance, PISE improves classification accuracy by 2.57% and reduces variance by 9x at 5% sampling.

Method: Created EgoPointVQA dataset with 4000 synthetic and 400 real-world videos across deictic reasoning tasks. Proposed Hand Intent Tokens (HINT) that encode tokens from 3D hand keypoints using off-the-shelf reconstruction model and interleave them with model input.

Result: HINT-14B achieves 68.1% accuracy on average over 6 tasks, surpassing state-of-the-art InternVL3-14B by 6.6%. The model outperforms others across different backbones and model sizes.

Conclusion: The work addresses a critical gap in MLLMs’ ability to understand pointing gestures in egocentric videos, providing a dataset, benchmark, and effective method (HINT) that significantly improves performance on gesture-grounded question answering tasks.

Abstract: Understanding and answering questions based on a user’s pointing gesture is essential for next-generation egocentric AI assistants. However, current Multimodal Large Language Models (MLLMs) struggle with such tasks due to the lack of gesture-rich data and their limited ability to infer fine-grained pointing intent from egocentric video. To address this, we introduce EgoPointVQA, a dataset and benchmark for gesture-grounded egocentric question answering, comprising 4000 synthetic and 400 real-world videos across multiple deictic reasoning tasks. Built upon it, we further propose Hand Intent Tokens (HINT), which encodes tokens derived from 3D hand keypoints using an off-the-shelf reconstruction model and interleaves them with the model input to provide explicit spatial and temporal context for interpreting pointing intent. We show that our model outperforms others in different backbones and model sizes. In particular, HINT-14B achieves 68.1% accuracy, on average over 6 tasks, surpassing the state-of-the-art, InternVL3-14B, by 6.6%. To further facilitate the open research, we will release the code, model, and dataset. Project page: https://yuuraa.github.io/papers/choi2026egovqa

Method: Proposes SERA with two complementary stages: 1) SERA-Adapter inserts expression-conditioned adapters into selected backbone blocks for expert-guided refinement and cross-modal attention, and 2) SERA-Fusion reshapes token features into spatial grids and applies geometry-preserving expert transformations before multimodal interaction. Includes lightweight routing mechanism and parameter-efficient tuning that updates less than 1% of backbone parameters.

Result: Experiments on standard referring image segmentation benchmarks show SERA consistently outperforms strong baselines, with especially clear gains on expressions requiring accurate spatial localization and precise boundary delineation.

Conclusion: SERA effectively addresses limitations of uniform refinement strategies in referring image segmentation by introducing expression-aware expert routing that improves spatial coherence and boundary precision while maintaining computational efficiency through parameter-efficient tuning.

Abstract: Referring image segmentation aims to produce a pixel-level mask for the image region described by a natural-language expression. Although pretrained vision-language models have improved semantic grounding, many existing methods still rely on uniform refinement strategies that do not fully match the diverse reasoning requirements of referring expressions. Because of this mismatch, predictions often contain fragmented regions, inaccurate boundaries, or even the wrong object, especially when pretrained backbones are frozen for computational efficiency. To address these limitations, we propose SERA, a Spatio-Semantic Expert Routing Architecture for referring image segmentation. SERA introduces lightweight, expression-aware expert refinement at two complementary stages within a vision-language framework. First, we design SERA-Adapter, which inserts an expression-conditioned adapter into selected backbone blocks to improve spatial coherence and boundary precision through expert-guided refinement and cross-modal attention. We then introduce SERA-Fusion, which strengthens intermediate visual representations by reshaping token features into spatial grids and applying geometry-preserving expert transformations before multimodal interaction. In addition, a lightweight routing mechanism adaptively weights expert contributions while remaining compatible with pretrained representations. To make this routing stable under frozen encoders, SERA uses a parameter-efficient tuning strategy that updates only normalization and bias terms, affecting less than 1% of the backbone parameters. Experiments on standard referring image segmentation benchmarks show that SERA consistently outperforms strong baselines, with especially clear gains on expressions that require accurate spatial localization and precise boundary delineation.

Method: Conducted a controlled diagnostic study within the LLaVA framework to isolate how design choices affect spatial grounding. Evaluated frontier models and LLaVA variants on spatial benchmarks, comparing CLIP-based encoders against alternatives trained with denser or generative objectives, and variants augmented with 2D positional encoding.

Result: Results show consistent spatial performance gaps across models. Encoder objectives and positional structure shape spatial behavior but do not fully resolve the spatial reasoning limitations.

Conclusion: Current VLM design choices (CLIP-style encoders and 1D tokenization with 1D positional encoding) contribute significantly to spatial reasoning limitations. While encoder objectives and positional structure affect spatial behavior, they don’t fully solve the problem, suggesting need for architectural innovations.

Abstract: Vision-language models (VLMs) have advanced rapidly, yet they still struggle with basic spatial reasoning. Despite strong performance on general benchmarks, modern VLMs remain brittle at understanding 2D spatial relationships such as relative position, layout, and counting. We argue that this failure is not merely a data problem, but is closely tied to dominant design choices in current VLM pipelines: reliance on CLIP-style image encoders and the flattening of images into 1D token sequences with 1D positional encoding. We present a controlled diagnostic study within the LLaVA framework to isolate how these choices affect spatial grounding. We evaluate frontier models and LLaVA variants on a suite of spatial benchmarks, comparing CLIP-based encoders against alternatives trained with denser or generative objectives, as well as variants augmented with 2D positional encoding. Our results show consistent spatial performance gaps across models, and indicate that encoder objectives and positional structure shape spatial behavior, but do not fully resolve it.

Method: Proposes Deco-Mamba with U-Net-like structure: encoder combines CNN block and Transformer backbone; decoder integrates Co-Attention Gate (CAG), Vision State Space Module (VSSM), and deformable convolutional refinement block. Uses windowed distribution-aware KL-divergence loss for deep supervision.

Result: Achieves state-of-the-art performance on diverse medical image segmentation benchmarks with strong generalization capability while maintaining moderate model complexity.

Conclusion: Decoder-centric approach with novel architectural components enables generalized medical image segmentation across diverse modalities with efficient computation.

Abstract: Deep learning has achieved remarkable success in medical image segmentation, often reaching expert-level accuracy in delineating tumors and tissues. However, most existing approaches remain task-specific, showing strong performance on individual datasets but limited generalization across diverse imaging modalities. Moreover, many methods focus primarily on the encoder, relying on large pretrained backbones that increase computational complexity. In this paper, we propose a decoder-centric approach for generalized 2D medical image segmentation. The proposed Deco-Mamba follows a U-Net-like structure with a Transformer-CNN-Mamba design. The encoder combines a CNN block and Transformer backbone for efficient feature extraction, while the decoder integrates our novel Co-Attention Gate (CAG), Vision State Space Module (VSSM), and deformable convolutional refinement block to enhance multi-scale contextual representation. Additionally, a windowed distribution-aware KL-divergence loss is introduced for deep supervision across multiple decoding stages. Extensive experiments on diverse medical image segmentation benchmarks yield state-of-the-art performance and strong generalization capability while maintaining moderate model complexity. The source code will be released upon acceptance.

Method: Proposes CLGT framework with three components: 1) Causal Feature Extractor (CFE) that uses causal intervention to focus on stable, task-relevant semantics; 2) Geometric Topology Fusion module that injects Bird’s Eye View road topology into street features; 3) Data-Adaptive Pooling module to enhance representation of semantically rich regions.

Result: Extensive experiments on CVUSA, CVACT, and their robustness-enhanced variants show CLGT achieves state-of-the-art performance, particularly under challenging real-world corruptions.

Conclusion: The CLGT framework effectively addresses cross-view inconsistencies and confounding factors in geo-localization through causal learning and geometric topology fusion, demonstrating superior robustness and performance.

Abstract: Cross-view geo-localization (CVGL) aims to estimate the geographic location of a street image by matching it with a corresponding aerial image. This is critical for autonomous navigation and mapping in complex real-world scenarios. However, the task remains challenging due to significant viewpoint differences and the influence of confounding factors. To tackle these issues, we propose the Causal Learning and Geometric Topology (CLGT) framework, which integrates two key components: a Causal Feature Extractor (CFE) that mitigates the influence of confounding factors by leveraging causal intervention to encourage the model to focus on stable, task-relevant semantics; and a Geometric Topology Fusion (GT Fusion) module that injects Bird’s Eye View (BEV) road topology into street features to alleviate cross-view inconsistencies caused by extreme perspective changes. Additionally, we introduce a Data-Adaptive Pooling (DA Pooling) module to enhance the representation of semantically rich regions. Extensive experiments on CVUSA, CVACT, and their robustness-enhanced variants (CVUSA-C-ALL and CVACT-C-ALL) demonstrate that CLGT achieves state-of-the-art performance, particularly under challenging real-world corruptions. Our codes are available at https://github.com/oyss-szu/CLGT.

Method: Proposes AccelAes, a training-free framework that builds AesMask (aesthetic focus mask) from prompt semantics and cross-attention signals. Uses SkipSparse to reallocate computation to masked regions when localized computation is feasible, and employs a lightweight step-level prediction cache to reduce temporal redundancy by periodically replacing full Transformer evaluations.

Result: Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, achieves 2.11× speedup and improves ImageReward by +11.9% over dense baseline.

Conclusion: AccelAes effectively accelerates DiTs through aesthetics-aware spatio-temporal reduction while enhancing perceptual aesthetics, offering a practical solution for deployment without requiring retraining.

Abstract: Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.

Method: Proposes DINOLight framework with: 1) adaptive feature fusion module combining DINOv2 layers using point-wise softmax mask, 2) restoration network integrating fused features in spatial and frequency domains via auxiliary cross-attention mechanism.

Result: Achieves superior performance on Ambient6K dataset; shows DINOv2 features effectively enhance ambient light normalization; achieves competitive results on shadow-removal benchmarks compared to methods using mask priors.

Conclusion: DINOLight successfully integrates DINOv2’s visual understanding capabilities for ambient light normalization, demonstrating the effectiveness of self-supervised vision features for lighting restoration tasks.

Abstract: This paper presents a new ambient light normalization framework, DINOLight, that integrates the self-supervised model DINOv2’s image understanding capability into the restoration process as a visual prior. Ambient light normalization aims to restore images degraded by non-uniform shadows and lighting caused by multiple light sources and complex scene geometries. We observe that DINOv2 can reliably extract both semantic and geometric information from a degraded image. Based on this observation, we develop a novel framework to utilize DINOv2 features for lighting normalization. First, we propose an adaptive feature fusion module that combines features from different DINOv2 layers using a point-wise softmax mask. Next, the fused features are integrated into our proposed restoration network in both spatial and frequency domains through an auxiliary cross-attention mechanism. Experiments show that DINOLight achieves superior performance on the Ambient6K dataset, and that DINOv2 features are effective for enhancing ambient light normalization. We also apply our method to shadow-removal benchmark datasets, achieving competitive results compared to methods that use mask priors. Codes will be released upon acceptance.

Method: MRGeo employs hierarchical defense: 1) Spatial-Channel Enhancement Block with Spatial Adaptive Representation Module (parallel global/local features with dynamic gating) and Channel Calibration Module (multi-granularity channel dependencies); 2) Region-level Geometric Alignment Module for coarse-grained spatial consistency.

Result: Achieves average R@1 improvement of 2.92% across three robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, CVACT_test-C-ALL) and superior performance in cross-area evaluation, demonstrating robustness and generalization.

Conclusion: MRGeo addresses the critical robustness gap in CVGL, providing systematic defense against image corruption while maintaining performance, enabling more reliable real-world deployment of geo-localization systems.

Abstract: Cross-view geo-localization (CVGL) aims to accurately localize street-view images through retrieval of corresponding geo-tagged satellite images. While prior works have achieved nearly perfect performance on certain standard datasets, their robustness in real-world corrupted environments remains under-explored. This oversight causes severe performance degradation or failure when images are affected by corruption such as blur or weather, significantly limiting practical deployment. To address this critical gap, we introduce MRGeo, the first systematic method designed for robust CVGL under corruption. MRGeo employs a hierarchical defense strategy that enhances the intrinsic quality of features and then enforces a robust geometric prior. Its core is the Spatial-Channel Enhancement Block, which contains: (1) a Spatial Adaptive Representation Module that models global and local features in parallel and uses a dynamic gating mechanism to arbitrate their fusion based on feature reliability; and (2) a Channel Calibration Module that performs compensatory adjustments by modeling multi-granularity channel dependencies to counteract information loss. To prevent spatial misalignment under severe corruption, a Region-level Geometric Alignment Module imposes a geometric structure on the final descriptors, ensuring coarse-grained consistency. Comprehensive experiments on both robustness benchmark and standard datasets demonstrate that MRGeo not only achieves an average R@1 improvement of 2.92% across three comprehensive robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) but also establishes superior performance in cross-area evaluation, thereby demonstrating its robustness and generalization capability.

Method: Built on a ViT backbone, SDF-Net introduces structure consistency constraints using scale-invariant gradient energy statistics from intermediate layers to anchor representations against radiometric variations. It disentangles learned representations into modality-invariant identity features and modality-specific characteristics, then integrates them through parameter-free additive residual fusion.

Result: Extensive experiments on the HOSS-ReID dataset demonstrate that SDF-Net consistently outperforms existing state-of-the-art methods.

Conclusion: SDF-Net effectively incorporates geometric consistency into optical-SAR ship ReID, addressing the fundamental challenge of radiometric discrepancy by leveraging the physical prior of stable geometric structures across sensing modalities.

Abstract: Cross-modal ship re-identification (ReID) between optical and synthetic aperture radar (SAR) imagery is fundamentally challenged by the severe radiometric discrepancy between passive optical imaging and coherent active radar sensing. While existing approaches primarily rely on statistical distribution alignment or semantic matching, they often overlook a critical physical prior: ships are rigid objects whose geometric structures remain stable across sensing modalities, whereas texture appearance is highly modality-dependent. In this work, we propose SDF-Net, a Structure-Aware Disentangled Feature Learning Network that systematically incorporates geometric consistency into optical–SAR ship ReID. Built upon a ViT backbone, SDF-Net introduces a structure consistency constraint that extracts scale-invariant gradient energy statistics from intermediate layers to robustly anchor representations against radiometric variations. At the terminal stage, SDF-Net disentangles the learned representations into modality-invariant identity features and modality-specific characteristics. These decoupled cues are then integrated through a parameter-free additive residual fusion, effectively enhancing discriminative power. Extensive experiments on the HOSS-ReID dataset demonstrate that SDF-Net consistently outperforms existing state-of-the-art methods. The code and trained models are publicly available at https://github.com/cfrfree/SDF-Net.

Method: Neural Gate uses neuron-level model editing: learns a feature vector to identify neurons associated with privacy concepts within subject representations, then precisely updates model parameters to enhance refusal behavior for privacy-related questions.

Result: The method significantly boosts privacy protection (increased refusal rates for privacy queries) while preserving original utility on standard tasks, demonstrated through comprehensive experiments on MiniGPT and LLaVA models.

Conclusion: Neural Gate provides an effective approach to mitigate privacy risks in LVLMs through targeted neuron editing, achieving better generalization to novel privacy queries and maintaining model performance.

Abstract: Large Vision-Language Models (LVLMs) have shown remarkable potential across a wide array of vision-language tasks, leading to their adoption in critical domains such as finance and healthcare. However, their growing deployment also introduces significant security and privacy risks. Malicious actors could potentially exploit these models to extract sensitive information, highlighting a critical vulnerability. Recent studies show that LVLMs often fail to consistently refuse instructions designed to compromise user privacy. While existing work on privacy protection has made meaningful progress in preventing the leakage of sensitive data, they are constrained by limitations in both generalization and non-destructiveness. They often struggle to robustly handle unseen privacy-related queries and may inadvertently degrade a model’s performance on standard tasks. To address these challenges, we introduce Neural Gate, a novel method for mitigating privacy risks through neuron-level model editing. Our method improves a model’s privacy safeguards by increasing its rate of refusal for privacy-related questions, crucially extending this protective behavior to novel sensitive queries not encountered during the editing process. Neural Gate operates by learning a feature vector to identify neurons associated with privacy-related concepts within the model’s representation of a subject. This localization then precisely guides the update of model parameters. Through comprehensive experiments on MiniGPT and LLaVA, we demonstrate that our method significantly boosts the model’s privacy protection while preserving its original utility.

Method: Proposes Prediction-As-Perception (PAP) framework with two modules: prediction module forecasts future positions of ego vehicles and traffic participants using current frame perception results, and perception module uses these predictions as queries for subsequent frame processing in an iterative feedback loop.

Result: When evaluated on UniAD model with nuScenes dataset, PAP improves target tracking accuracy by 10% and increases inference speed by 15%, demonstrating enhanced efficiency and reduced computational resource consumption.

Conclusion: The biomimetic PAP framework successfully enhances perception model performance by integrating prediction capabilities, showing significant improvements in both accuracy and efficiency for 3D object perception tasks.

Abstract: Humans combine prediction and perception to observe the world. When faced with rapidly moving birds or insects, we can only perceive them clearly by predicting their next position and focusing our gaze there. Inspired by this, this paper proposes the Prediction-As-Perception (PAP) framework, integrating a prediction-perception architecture into 3D object perception tasks to enhance the model’s perceptual accuracy. The PAP framework consists of two main modules: prediction and perception, primarily utilizing continuous frame information as input. Firstly, the prediction module forecasts the potential future positions of ego vehicles and surrounding traffic participants based on the perception results of the current frame. These predicted positions are then passed as queries to the perception module of the subsequent frame. The perceived results are iteratively fed back into the prediction module. We evaluated the PAP structure using the end-to-end model UniAD on the nuScenes dataset. The results demonstrate that the PAP structure improves UniAD’s target tracking accuracy by 10% and increases the inference speed by 15%. This indicates that such a biomimetic design significantly enhances the efficiency and accuracy of perception models while reducing computational resource consumption.

Method: Created A2Z dataset with 10M annotations for 1M ABC CAD models including high-res meshes, 3D sketches, BRep geometric/topological data, and text captions. Added 25K professional CAD models. Trained foundation model on 150K subset for BRep co-edge and corner detection from 3D scans.

Result: Largest compilation of multi-modal CAD annotations (5TB), assessed with novel metrics, GPT-5, Gemini, and human feedback. Foundation model benchmarked for CAD reverse engineering tasks.

Conclusion: A2Z dataset enables unprecedented BRep learning for CAD reverse engineering and retrieval. Dataset, metrics, and checkpoints will be publicly released to support research.

Abstract: Reverse engineering and rapid prototyping of computer-aided design (CAD) models from 3D scans, sketches, or simple text prompts are vital in industrial product design. However, recent advances in geometric deep learning techniques lack a multi-modal understanding of parametric CAD features stored in their boundary representation (BRep). This study presents the largest compilation of 10 million multi-modal annotations and metadata for 1 million ABC CAD models, namely A2Z, to unlock an unprecedented level of BRep learning. A2Z comprises (i) high-resolution meshes with salient 3D scanning features, (ii) 3D hand-drawn sketches equipped with (iii) geometric and topological information about BRep co-edges, corners, and surfaces, and (iv) textual captions and tags describing the product in the mechanical world. Creating such carefully structured, large-scale data, which requires nearly 5 terabytes of storage to leverage unparalleled CAD learning/retrieval tasks, is very challenging. The scale, quality, and diversity of our multi-modal annotations are assessed using novel metrics, GPT-5, Gemini, and extensive human feedback mechanisms. To this end, we also merge an additional 25,000 CAD models of electronic enclosures (e.g., tablets, ports) designed by skilled professionals with our A2Z dataset. Subsequently, we train and benchmark a foundation model on a subset of 150K CAD models to detect BRep co-edges and corner vertices from 3D scans, a key downstream task in CAD reverse engineering. The annotated dataset, metrics, and checkpoints will be publicly released to support numerous research directions.

Method: Introduces D-Negation dataset with objects annotated with positive/negative semantic descriptions, and a grouped opposition-based learning framework with complementary loss functions for negation reasoning.

Result: Improvements of up to 4.4 mAP and 5.7 mAP on positive/negative semantic evaluations by fine-tuning <10% of model parameters in state-of-the-art grounding model.

Conclusion: Explicitly modeling negation semantics substantially enhances robustness and localization accuracy of vision-language grounding models.

Abstract: Current vision-language detection and grounding models predominantly focus on prompts with positive semantics and often struggle to accurately interpret and ground complex expressions containing negative semantics. A key reason for this limitation is the lack of high-quality training data that explicitly captures discriminative negative samples and negation-aware language descriptions. To address this challenge, we introduce D-Negation, a new dataset that provides objects annotated with both positive and negative semantic descriptions. Building upon the observation that negation reasoning frequently appears in natural language, we further propose a grouped opposition-based learning framework that learns negation-aware representations from limited samples. Specifically, our method organizes opposing semantic descriptions from D-Negation into structured groups and formulates two complementary loss functions that encourage the model to reason about negation and semantic qualifiers. We integrate the proposed dataset and learning strategy into a state-of-the-art language-based grounding model. By fine-tuning fewer than 10 percent of the model parameters, our approach achieves improvements of up to 4.4 mAP and 5.7 mAP on positive and negative semantic evaluations, respectively. These results demonstrate that explicitly modeling negation semantics can substantially enhance the robustness and localization accuracy of vision-language grounding models.

Method: Introduces RoboStereo, a symmetric dual-tower 4D world model with bidirectional cross-modal enhancement for spatiotemporal geometric consistency. Presents a unified framework with three components: Test-Time Policy Augmentation (TTPA) for pre-execution verification, Imitative-Evolutionary Policy Learning (IEPL) using visual perceptual rewards from expert demonstrations, and Open-Exploration Policy Learning (OEPL) for autonomous skill discovery.

Result: RoboStereo achieves state-of-the-art generation quality, and the unified framework delivers >97% average relative improvement on fine-grained manipulation tasks.

Conclusion: RoboStereo provides a high-fidelity 4D simulator that addresses geometric consistency issues in embodied world models and enables effective policy optimization through a comprehensive framework combining verification, imitation learning, and autonomous exploration.

Abstract: Scalable Embodied AI faces fundamental constraints due to prohibitive costs and safety risks of real-world interaction. While Embodied World Models (EWMs) offer promise through imagined rollouts, existing approaches suffer from geometric hallucinations and lack unified optimization frameworks for practical policy improvement. We introduce RoboStereo, a symmetric dual-tower 4D world model that employs bidirectional cross-modal enhancement to ensure spatiotemporal geometric consistency and alleviate physics hallucinations. Building upon this high-fidelity 4D simulator, we present the first unified framework for world-model-based policy optimization: (1) Test-Time Policy Augmentation (TTPA) for pre-execution verification, (2) Imitative-Evolutionary Policy Learning (IEPL) leveraging visual perceptual rewards to learn from expert demonstrations, and (3) Open-Exploration Policy Learning (OEPL) enabling autonomous skill discovery and self-correction. Comprehensive experiments demonstrate RoboStereo achieves state-of-the-art generation quality, with our unified framework delivering >97% average relative improvement on fine-grained manipulation tasks.

Method: Proposes a structure-aware Salient Gaussian representation initialized from geometric and reflectance feature points from LiDAR, refined through salient transform and improved density control. Calibrates LiDAR intensity into reflectance attached to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB for boundary consistency.

Result: Extensive experiments on Waymo Open Dataset show superior reconstruction performance with fewer Gaussians and shorter training time. Specifically, on Complex Lighting scenes, surpasses OmniRe by 1.18 dB PSNR.

Conclusion: LR-SGS effectively leverages LiDAR reflectance and geometric features to create a robust and efficient Gaussian splatting method for self-driving scenes, particularly excelling in challenging lighting conditions.

Abstract: Recent 3D Gaussian Splatting (3DGS) methods have demonstrated the feasibility of self-driving scene reconstruction and novel view synthesis. However, most existing methods either rely solely on cameras or use LiDAR only for Gaussian initialization or depth supervision, while the rich scene information contained in point clouds, such as reflectance, and the complementarity between LiDAR and RGB have not been fully exploited, leading to degradation in challenging self-driving scenes, such as those with high ego-motion and complex lighting. To address these issues, we propose a robust and efficient LiDAR-reflectance-guided Salient Gaussian Splatting method (LR-SGS) for self-driving scenes, which introduces a structure-aware Salient Gaussian representation, initialized from geometric and reflectance feature points extracted from LiDAR and refined through a salient transform and improved density control to capture edge and planar structures. Furthermore, we calibrate LiDAR intensity into reflectance and attach it to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB to enforce boundary consistency. Extensive experiments on the Waymo Open Dataset demonstrate that LR-SGS achieves superior reconstruction performance with fewer Gaussians and shorter training time. In particular, on Complex Lighting scenes, our method surpasses OmniRe by 1.18 dB PSNR.

Method: Proposes Multi-View GRPO (MV-GRPO) that augments condition space using a flexible Condition Enhancer to generate semantically adjacent yet diverse captions for each prompt, enabling multi-view advantage re-estimation without costly sample regeneration.

Result: Extensive experiments demonstrate MV-GRPO achieves superior alignment performance over state-of-the-art methods in text-to-image preference alignment.

Conclusion: MV-GRPO effectively addresses the limitations of single-view evaluation in GRPO by creating dense multi-view reward mappings, leading to better exploration of inter-sample relationships and improved alignment performance.

Abstract: Group Relative Policy Optimization (GRPO) has emerged as a powerful framework for preference alignment in text-to-image (T2I) flow models. However, we observe that the standard paradigm where evaluating a group of generated samples against a single condition suffers from insufficient exploration of inter-sample relationships, constraining both alignment efficacy and performance ceilings. To address this sparse single-view evaluation scheme, we propose Multi-View GRPO (MV-GRPO), a novel approach that enhances relationship exploration by augmenting the condition space to create a dense multi-view reward mapping. Specifically, for a group of samples generated from one prompt, MV-GRPO leverages a flexible Condition Enhancer to generate semantically adjacent yet diverse captions. These captions enable multi-view advantage re-estimation, capturing diverse semantic attributes and providing richer optimization signals. By deriving the probability distribution of the original samples conditioned on these new captions, we can incorporate them into the training process without costly sample regeneration. Extensive experiments demonstrate that MV-GRPO achieves superior alignment performance over state-of-the-art methods.

Method: Repurpose latent tokens of frozen VGGT as world state, train lightweight temporal flow transformer to autoregressively predict future trajectory. Address high-dimensional feature space challenges with clean-target parameterization and two-stage latent flow-forcing curriculum.

Result: Significantly outperforms baselines in depth forecasting on KITTI, Cityscapes, and TartanAir, runs 3.6-5 times faster with only 0.43B trainable parameters.

Conclusion: Frozen geometry-foundation-model features serve as effective and efficient predictive state for 3D world modeling, offering better geometric consistency than video-based approaches.

Abstract: World models that forecast scene evolution by generating future video frames devote the bulk of their capacity to photometric details, yet the resulting predictions often remain geometrically inconsistent. We present VGGT-World, a geometry world model that side-steps video generation entirely and instead forecasts the temporal evolution of frozen geometry-foundation-model (GFM) features. Concretely, we repurpose the latent tokens of a frozen VGGT as the world state and train a lightweight temporal flow transformer to autoregressively predict their future trajectory. Two technical challenges arise in this high-dimensional (d=1024) feature space: (i) standard velocity-prediction flow matching collapses, and (ii) autoregressive rollout suffers from compounding exposure bias. We address the first with a clean-target (z-prediction) parameterization that yields a substantially higher signal-to-noise ratio, and the second with a two-stage latent flow-forcing curriculum that progressively conditions the model on its own partially denoised rollouts. Experiments on KITTI, Cityscapes, and TartanAir demonstrate that VGGT-World significantly outperforms the strongest baselines in depth forecasting while running 3.6-5 times faster with only 0.43B trainable parameters, establishing frozen GFM features as an effective and efficient predictive state for 3D world modeling.

Method: Introduces lightweight scale alignment stage to restore multi-view scale coherence, then integrates pretrained VFM features via task-specific adapters trained for reconstruction while preserving cross-domain robustness of pretrained representations.

Result: Achieves state-of-the-art performance across ScanNet test split, TUM RGB-D, and Tanks and Temples datasets. On challenging outdoor Tanks and Temples, achieves F1 score of 70.1 vs 51.8 for closest competitor VGGT.

Conclusion: Successfully bridges transferable VFM priors with scale-consistent requirements for scene-level neural reconstruction, demonstrating robust performance across domains including challenging outdoor scenes.

Abstract: Scene-level neural volumetric reconstruction from monocular videos remains challenging, especially under severe domain shifts. Although recent advances in vision foundation models (VFMs) provide transferable generalized priors learned from large-scale data, their scaleambiguous predictions are incompatible with the scale consistency required by volumetric fusion. To address this gap, we present VFMRecon, the first attempt to bridge transferable VFM priors with scaleconsistent requirements in scene-level neural reconstruction. Specifically, we first introduce a lightweight scale alignment stage that restores multiview scale coherence. We then integrate pretrained VFM features into the neural volumetric reconstruction pipeline via lightweight task-specific adapters, which are trained for reconstruction while preserving the crossdomain robustness of pretrained representations. We train our model on ScanNet train split and evaluate on both in-distribution ScanNet test split and out-of-distribution TUM RGB-D and Tanks and Temples datasets. The results demonstrate that our model achieves state-of-theart performance across all datasets domains. In particular, on the challenging outdoor Tanks and Temples dataset, our model achieves an F1 score of 70.1 in reconstructed mesh evaluation, substantially outperforming the closest competitor, VGGT, which only attains 51.8.

Method: AVION uses a teacher-student knowledge distillation framework: the teacher constructs semantically rich textual prototypes using LLM-generated descriptions verified by remote sensing image features; the student integrates lightweight learnable prompts into both vision and language encoders, guided by the teacher to align embeddings and cross-modal relationships.

Result: Experiments on six optical remote sensing benchmarks show improved few-shot classification and base-class accuracy without degrading generalization to novel categories, enhanced mean recall for cross-modal retrieval, with minimal additional trainable parameters.

Conclusion: AVION effectively adapts vision-language models to remote sensing imagery through knowledge distillation, addressing semantic coverage and feature adaptability issues while maintaining efficiency.

Abstract: Adapting vision-language models to remote sensing imagery remains challenging due to two key factors: limited semantic coverage in textual representations and insufficient adaptability of visual features. These issues are particularly significant in aerial scenes, which involve various visual appearances and fine-grained object distinctions. We propose AVION, a knowledge distillation framework tailored for remote sensing adaptation of vision-language models. The teacher module constructs semantically rich textual prototypes by collecting descriptions from a large language model and verifying validity using remote sensing image features. The student module integrates lightweight and learnable prompts into both vision and language encoders, guided by the teacher to align embeddings and their cross-modal relationships. Once trained, the student operates independently during inference. Experiments on six optical remote sensing benchmarks show that AVION improves few-shot classification and base-class accuracy without degrading generalization to novel categories. It also enhances mean recall for cross-modal retrieval, with minimal additional trainable parameters.

Method: 1) Constructed Multi-modal Panoramic 3D Outdoor (MPO) dataset with point clouds from two different LiDARs labeled with six categories. 2) Developed CNNs that take omnidirectional depth/reflectance images as inputs for outdoor place categorization. 3) Used both depth and reflectance modalities and visualized learned features for analysis.

Result: The approach outperforms traditional methods on the MPO dataset, demonstrating effectiveness of using both depth and reflectance modalities. Feature visualizations provide insights into what the networks learn.

Conclusion: The proposed CNN-based method using LiDAR depth/reflectance data effectively categorizes outdoor places, addressing challenges of perceptual variations in outdoor environments for autonomous systems.

Abstract: Semantic place categorization, which is one of the essential tasks for autonomous robots and vehicles, allows them to have capabilities of self-decision and navigation in unfamiliar environments. In particular, outdoor places are more difficult targets than indoor ones due to perceptual variations, such as dynamic illuminance over twenty-four hours and occlusions by cars and pedestrians. This paper presents a novel method of categorizing outdoor places using convolutional neural networks (CNNs), which take omnidirectional depth/reflectance images obtained by 3D LiDARs as the inputs. First, we construct a large-scale outdoor place dataset named Multi-modal Panoramic 3D Outdoor (MPO) comprising two types of point clouds captured by two different LiDARs. They are labeled with six outdoor place categories: coast, forest, indoor/outdoor parking, residential area, and urban area. Second, we provide CNNs for LiDAR-based outdoor place categorization and evaluate our approach with the MPO dataset. Our results on the MPO dataset outperform traditional approaches and show the effectiveness in which we use both depth and reflectance modalities. To analyze our trained deep networks we visualize the learned features.

Method: Introduces focal error diversity to capture complementary reasoning across VLMs and CKA-focal metric to measure visual embedding disagreements. Uses Genetic Algorithm to prune ineffective VLMs from ensemble surface, then fuses outputs of selected models to produce dual focal-diversity predictions.

Result: Outperforms best-performing VLM on MMMU by 8.09% and MMMU-Pro by 4.87% accuracy. For generative tasks, beats top-2 VLM performers (Intern-VL2-8b and Qwen2.5-VL-7b) on A-OKVQA and OCR-VQA benchmarks.

Conclusion: V3Fusion effectively combines heterogeneous VLMs using both vision and language modalities, dynamically captures epistemic uncertainty, mitigates hallucinations, and achieves superior performance even when majority of VLMs make incorrect predictions.

Abstract: With the growing number and diversity of Vision-Language Models (VLMs), many works explore language-based ensemble, collaboration, and routing techniques across multiple VLMs to improve multi-model reasoning. In contrast, we address the diverse model selection using both vision and language modalities. We introduce focal error diversity to capture complementary reasoning across VLMs and a CKA-based focal diversity metric (CKA-focal) to measure disagreement in their visual embeddings. On the constructed ensemble surface from a pool of candidate VLMs, we applied a Genetic Algorithm to effectively prune out those component VLMs that do not add value to the fusion performance. We identify the best combination for each task as well as fuse the outputs of each VLMs in the model pool, and show that heterogeneous models can capture epistemic uncertainty dynamically and mitigate hallucinations. Our V3Fusion approach is capable of producing dual focal-diversity fused predictions with high performance for vision-language reasoning, even when there is no majority consensus or the majority of VLMs make incorrect predictions. Extensive experiments validate V3Fusion on four popular VLM benchmarks (A-OKVQA, MMMU, MMMU-Pro, and OCR-VQA). The results show that V3Fusion outperforms the best-performing VLM on MMMU by 8.09% and MMMU-Pro by 4.87% gain in accuracy. For generative tasks, V3Fusion outperforms Intern-VL2-8b and Qwen2.5-VL-7b, the top-2 VLM performers on both A-OKVQA and OCR-VQA. Our code and datasets are available at https://github.com/sftekin/v3fusion.

Method: Proposes GeoGran-Aware Hierarchical Feature Fusion Network (G2HFNet) with Swin Transformer backbone and four key modules: multi-scale detail enhancement (MDE) for scale variations, dual-branch geo-gran complementary (DGC) for fine-grained details and positional information, deep semantic perception (DSP) for refining high-level positional cues, and local-global guidance fusion (LGF) for multi-level feature integration.

Result: Extensive experiments demonstrate that G2HFNet achieves high-quality saliency maps and significantly improves detection performance in challenging remote sensing scenarios.

Conclusion: G2HFNet effectively addresses scale variations and complex backgrounds in remote sensing images through geometric and granular-aware hierarchical feature fusion, outperforming existing methods.

Abstract: Remote sensing images captured from aerial perspectives often exhibit significant scale variations and complex backgrounds, posing challenges for salient object detection (SOD). Existing methods typically extract multi-level features at a single scale using uniform attention mechanisms, leading to suboptimal representations and incomplete detection results. To address these issues, we propose a GeoGran-Aware Hierarchical Feature Fusion Network (G2HFNet) that fully exploits geometric and granular cues in optical remote sensing images. Specifically, G2HFNet adopts Swin Transformer as the backbone to extract multi-level features and integrates three key modules: the multi-scale detail enhancement (MDE) module to handle object scale variations and enrich fine details, the dual-branch geo-gran complementary (DGC) module to jointly capture fine-grained details and positional information in mid-level features, and the deep semantic perception (DSP) module to refine high-level positional cues via self-attention. Additionally, a local-global guidance fusion (LGF) module is introduced to replace traditional convolutions for effective multi-level feature integration. Extensive experiments demonstrate that G2HFNet achieves high-quality saliency maps and significantly improves detection performance in challenging remote sensing scenarios.

Method: MoKus uses a two-stage approach: (1) visual concept learning to learn anchor representations storing visual information, and (2) textual knowledge updating that updates answers to knowledge queries with anchor representations, enabling cross-modal knowledge transfer from text to visual generation.

Result: MoKus outperforms state-of-the-art methods on the new KnowCusBench benchmark for knowledge-aware concept customization. The framework also demonstrates extensions to virtual concept creation, concept erasure, and improvements on world knowledge benchmarks.

Conclusion: The proposed knowledge-aware concept customization task and MoKus framework effectively address limitations of existing methods by enabling cross-modal knowledge transfer, achieving superior performance and broader applicability to various knowledge-aware applications.

Abstract: Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.

Method: Two-stage approach: 1) Region guidance stage with three parallel branches using context interaction and spatial-aware fusion modules to generate guidance maps and similarity scores; 2) Saliency generation stage with selective optimization module for modality fusion, dense detail enhancement module for low-level features, and mutual interaction semantic module for high-level features.

Result: Achieves competitive performance against 27 state-of-the-art SOD methods on RGB-T datasets.

Conclusion: RSONet effectively addresses modality inconsistency in RGB-T salient object detection through region-guided selective optimization and comprehensive feature enhancement.

Abstract: This paper focuses on the inconsistency in salient regions between RGB and thermal images. To address this issue, we propose the Region-guided Selective Optimization Network for RGB-T Salient Object Detection, which consists of the region guidance stage and saliency generation stage. In the region guidance stage, three parallel branches with same encoder-decoder structure equipped with the context interaction (CI) module and spatial-aware fusion (SF) module are designed to generate the guidance maps which are leveraged to calculate similarity scores. Then, in the saliency generation stage, the selective optimization (SO) module fuses RGB and thermal features based on the previously obtained similarity values to mitigate the impact of inconsistent distribution of salient targets between the two modalities. After that, to generate high-quality detection result, the dense detail enhancement (DDE) module which adopts the multiple dense connections and visual state space blocks is applied to low-level features for optimizing the detail information. In addition, the mutual interaction semantic (MIS) module is placed in the high-level features to dig the location cues by the mutual fusion strategy. We conduct extensive experiments on the RGB-T dataset, and the results demonstrate that the proposed RSONet achieves competitive performance against 27 state-of-the-art SOD methods.

Method: STRAP-ViT uses Jensen-Shannon Divergence to detect tokens with anomalous statistical properties (those overlapping with adversarial patches) during a Detection Phase, then applies randomized composite transformations on these tokens during a Mitigation Phase. It’s a non-trainable plug-and-play block that transforms enough tokens to cover at least 50% of the adversarial patch.

Result: The method achieves robust accuracies within 2-3% of clean baselines across multiple pre-trained ViT architectures (ViT-base-16, DinoV2), datasets (ImageNet, CalTech-101), and adversarial attacks (Adversarial Patch, LAVAN, GDPA, RP2), outperforming state-of-the-art defenses.

Conclusion: STRAP-ViT provides an effective, computationally efficient defense against adversarial patches for Vision Transformers without requiring retraining, demonstrating strong robustness across diverse architectures and attack methods.

Abstract: Adversarial patches are physically realizable localized noise, which are able to hijack Vision Transformers (ViT) self-attention, pulling focus toward a small, high-contrast region and corrupting the class token to force confident misclassifications. In this paper, we claim that the tokens which correspond to the areas of the image that contain the adversarial noise, have different statistical properties when compared to the tokens which do not overlap with the adversarial perturbations. We use this insight to propose a mechanism, called STRAP-ViT, which uses Jensen-Shannon Divergence as a metric for segregating tokens that behave as anomalies in the Detection Phase, and then apply randomized composite transformations on them during the Mitigation Phase to make the adversarial noise ineffective. The minimum number of tokens to transform is a hyper-parameter for the defense mechanism and is chosen such that at least 50% of the patch is covered by the transformed tokens. STRAP-ViT fits as a non-trainable plug-and-play block within the ViT architectures, for inference purposes only, with a minimal computational cost and does not require any additional training cost/effort. STRAP-ViT has been tested on multiple pre-trained vision transformer architectures (ViT-base-16 and DinoV2) and datasets (ImageNet and CalTech-101), across multiple adversarial attacks (Adversarial Patch, LAVAN, GDPA and RP2), and found to provide excellent robust accuracies lying within a 2-3% range of the clean baselines, and outperform the state-of-the-art.

Method: Uses pre-trained EfficientNet-based emotion recognition models for facial embedding extraction. If model confidence exceeds threshold, uses its prediction; otherwise feeds embeddings into MLP trained on AffWild2. Applies sliding window smoothing to frame predictions. For violence detection, examines pre-trained architectures for frame embeddings and their aggregation for video classification.

Result: Experimental results on four ABAW challenge tasks demonstrate significant improvement in validation metrics over existing baselines.

Conclusion: The proposed approach effectively addresses multiple facial emotion understanding tasks and violence detection, showing competitive performance in the ABAW competition.

Abstract: This article presents our results for the 10th Affective Behavior Analysis in-the-Wild (ABAW) competition. For frame-wise facial emotion understanding tasks (frame-wise facial expression recognition, valence-arousal estimation, action unit detection), we propose a fast approach based on facial embedding extraction with pre-trained EfficientNet-based emotion recognition models. If the latter model’s confidence exceeds a threshold, its prediction is used. Otherwise, we feed embeddings into a simple multi-layered perceptron trained on the AffWild2 dataset. Estimated class-level scores are smoothed in a sliding window of fixed size to mitigate noise in frame-wise predictions. For the fine-grained violence detection task, we examine several pre-trained architectures for frame embeddings and their aggregation for video classification. Experimental results on four tasks from the ABAW challenge demonstrate that our approach significantly improves validation metrics over existing baselines.

Method: 1) Survey and categorize 30 feature matching algorithms (traditional and deep learning) into sparse, semidense, and dense methods. 2) Create CM-Bench with standardized evaluation across homography estimation, relative pose estimation, and geo-localization tasks. 3) Introduce adaptive preprocessing front-end that automatically selects enhancement strategies. 4) Present new IR-satellite dataset with manual ground-truth correspondences.

Result: Provides comprehensive evaluation of cross-modal feature matching methods, identifies performance gaps, and establishes standardized benchmark for future research. The new IR-satellite dataset enables practical geo-localization evaluation.

Conclusion: CM-Bench fills a critical gap in cross-modal feature matching research by providing standardized evaluation framework, comprehensive algorithm assessment, and new datasets, which will facilitate advancement in cross-modal visual perception applications.

Abstract: Infrared-visible (IR-VIS) feature matching plays an essential role in cross-modality visual localization, navigation and perception. Along with the rapid development of deep learning techniques, a number of representative image matching methods have been proposed. However, crossmodal feature matching is still a challenging task due to the significant appearance difference. A significant gap for cross-modal feature matching research lies in the absence of standardized benchmarks and metrics for evaluations. In this paper, we introduce a comprehensive cross-modal feature matching benchmark, CM-Bench, which encompasses 30 feature matching algorithms across diverse cross-modal datasets. Specifically, state-of-the-art traditional and deep learning-based methods are first summarized and categorized into sparse, semidense, and dense methods. These methods are evaluated by different tasks including homography estimation, relative pose estimation, and feature-matching-based geo-localization. In addition, we introduce a classification-network-based adaptive preprocessing front-end that automatically selects suitable enhancement strategies before matching. We also present a novel infrared-satellite cross-modal dataset with manually annotated ground-truth correspondences for practical geo-localization evaluation. The dataset and resource will be available at: https://github.com/SLZ98/CM-Bench.

Method: Propose VCBench with 406 videos containing frame-by-frame annotations of 10,071 event occurrence moments and object state changes. Create 8 fine-grained subcategories by decomposing world state maintenance into object counting (visible vs. cumulative unique identities) and event counting (instantaneous actions vs. complete activity cycles). Generate 1,000 streaming QA pairs with 4,576 query points along timelines.

Result: Evaluation on mainstream video-language models shows significant deficiencies in spatial-temporal state maintenance, particularly with tasks like periodic event counting. The benchmark provides three complementary metrics for diagnosing numerical precision, trajectory consistency, and temporal awareness.

Conclusion: VCBench offers a diagnostic framework for measuring and improving state maintenance in video understanding systems, revealing current models’ limitations in tracking world state over time.

Abstract: Video understanding requires models to continuously track and update world state during playback. While existing benchmarks have advanced video understanding evaluation across multiple dimensions, the observation of how models maintain world state remains insufficient. We propose VCBench, a streaming counting benchmark that repositions counting as a minimal probe for diagnosing world state maintenance capability. We decompose this capability into object counting (tracking currently visible objects vs.\ tracking cumulative unique identities) and event counting (detecting instantaneous actions vs.\ tracking complete activity cycles), forming 8 fine-grained subcategories. VCBench contains 406 videos with frame-by-frame annotations of 10,071 event occurrence moments and object state change moments, generating 1,000 streaming QA pairs with 4,576 query points along timelines. By observing state maintenance trajectories through streaming multi-point queries, we design three complementary metrics to diagnose numerical precision, trajectory consistency, and temporal awareness. Evaluation on mainstream video-language models shows that current models still exhibit significant deficiencies in spatial-temporal state maintenance, particularly struggling with tasks like periodic event counting. VCBench provides a diagnostic framework for measuring and improving state maintenance in video understanding systems.

Method: Proposes HFP-SAM with three key components: 1) Frequency Guided Adapter (FGA) to inject marine scene information into frozen SAM backbone using frequency domain prior masks, 2) Frequency-aware Point Selection (FPS) to generate highlighted regions through frequency analysis for point prompts, and 3) Full-View Mamba (FVM) to efficiently extract spatial and channel contextual information with linear complexity.

Result: Extensive experiments on four public datasets demonstrate superior performance compared to existing methods, showing effectiveness in marine animal segmentation tasks.

Conclusion: HFP-SAM effectively addresses the limitations of SAM for marine animal segmentation by incorporating frequency domain analysis and efficient contextual modeling, achieving state-of-the-art performance on benchmark datasets.

Abstract: Marine Animal Segmentation (MAS) aims at identifying and segmenting marine animals from complex marine environments. Most of previous deep learning-based MAS methods struggle with the long-distance modeling issue. Recently, Segment Anything Model (SAM) has gained popularity in general image segmentation. However, it lacks of perceiving fine-grained details and frequency information. To this end, we propose a novel learning framework, named Hierarchical Frequency Prompted SAM (HFP-SAM) for high-performance MAS. First, we design a Frequency Guided Adapter (FGA) to efficiently inject marine scene information into the frozen SAM backbone through frequency domain prior masks. Additionally, we introduce a Frequency-aware Point Selection (FPS) to generate highlighted regions through frequency analysis. These regions are combined with the coarse predictions of SAM to generate point prompts and integrate into SAM’s decoder for fine predictions. Finally, to obtain comprehensive segmentation masks, we introduce a Full-View Mamba (FVM) to efficiently extract spatial and channel contextual information with linear computational complexity. Extensive experiments on four public datasets demonstrate the superior performance of our approach. The source code is publicly available at https://github.com/Drchip61/TIP-HFP-SAM.

Method: IGASA uses a Hierarchical Pyramid Architecture (HPA) with two key components: Hierarchical Cross-Layer Attention (HCLA) module for multi-resolution feature alignment and local geometric consistency, and Iterative Geometry-Aware Refinement (IGAR) module for fine matching using reliable correspondences from coarse matching.

Result: IGASA significantly surpasses state-of-the-art methods on benchmark datasets (3D(Lo)Match, KITTI, nuScenes), achieving notable improvements in registration accuracy.

Conclusion: IGASA provides a robust foundation for advancing point cloud registration techniques and offers valuable insights for practical 3D vision applications.

Abstract: Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in \href{https://github.com/DongXu-Zhang/IGASA}{https://github.com/DongXu-Zhang/IGASA}.

Method: Proposes Text-Phase Synergy Network with Dual Priors (TPSNet): 1) Uses CLIP to generate domain-specific prompts as text priors for precise semantic supervision, 2) Introduces domain-invariant phase features as phase priors integrated into image representations to bridge domain gaps while preserving semantics.

Result: TPSNet significantly outperforms state-of-the-art methods on unsupervised cross-domain image retrieval benchmarks.

Conclusion: The synergy of text priors (CLIP-generated domain prompts) and phase priors (domain-invariant phase features) effectively addresses limitations of pseudo-label approaches and improves cross-domain retrieval performance.

Abstract: This paper studies unsupervised cross-domain image retrieval (UCDIR), which aims to retrieve images of the same category across different domains without relying on labeled data. Existing methods typically utilize pseudo-labels, derived from clustering algorithms, as supervisory signals for intra-domain representation learning and cross-domain feature alignment. However, these discrete pseudo-labels often fail to provide accurate and comprehensive semantic guidance. Moreover, the alignment process frequently overlooks the entanglement between domain-specific and semantic information, leading to semantic degradation in the learned representations and ultimately impairing retrieval performance. This paper addresses the limitations by proposing a Text-Phase Synergy Network with Dual Priors(TPSNet). Specifically, we first employ CLIP to generate a set of class-specific prompts per domain, termed as domain prompt, serving as a text prior that offers more precise semantic supervision. In parallel, we further introduce a phase prior, represented by domain-invariant phase features, which is integrated into the original image representations to bridge the domain distribution gaps while preserving semantic integrity. Leveraging the synergy of these dual priors, TPSNet significantly outperforms state-of-the-art methods on UCDIR benchmarks.

Method: Proposes CMHANet, a Cross-Modal Hybrid Attention Network that integrates 2D image contextual information with 3D point cloud geometric details. Uses contrastive learning-based optimization function to enforce geometric consistency and improve robustness to noise and partial observations.

Result: Achieves substantial improvements in registration accuracy and robustness on 3DMatch and 3DLoMatch datasets. Zero-shot evaluations on TUM RGB-D SLAM dataset verify generalization capability to unseen domains. Outperforms current techniques.

Conclusion: CMHANet effectively addresses robustness challenges in point cloud registration through cross-modal fusion and contrastive learning, demonstrating superior performance and generalization capabilities.

Abstract: Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model’s robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model’s generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in \href{https://github.com/DongXu-Zhang/CMHANet}{https://github.com/DongXu-Zhang/CMHANet}.

Method: Introduces Structural Semantic Unit (SSU) for relational semantic labeling and Coverage, Overlap, Trespass, and Excess (COTe) score as a decomposable metric for page parsing quality evaluation.

Result: COTe score is more informative than traditional metrics, reveals distinct failure modes, reduces interpretation-performance gap by up to 76% relative to F1, and works well even without explicit SSU labeling.

Conclusion: The proposed framework enables more robust, comparable, and nuanced DLA evaluation, with released SSU-labeled dataset and Python library for practical adoption.

Abstract: Document Layout analysis (DLA), is the process by which a page is parsed into meaningful elements, often using machine learning models. Typically, the quality of a model is judged using general object detection metrics such as IoU, F1 or mAP. However, these metrics are designed for images that are 2D projections of 3D space, not for the natively 2D imagery of printed media. This discrepancy can result in misleading or uninformative interpretation of model performance by the metrics. To encourage more robust, comparable, and nuanced DLA, we introduce: The Structural Semantic Unit (SSU) a relational labelling approach that shifts the focus from the physical to the semantic structure of the content; and the Coverage, Overlap, Trespass, and Excess (COTe) score, a decomposable metric for measuring page parsing quality. We demonstrate the value of these methods through case studies and by evaluating 5 common DLA models on 3 DLA datasets. We show that the COTe score is more informative than traditional metrics and reveals distinct failure modes across models, such as breaching semantic boundaries or repeatedly parsing the same region. In addition, the COTe score reduces the interpretation-performance gap by up to 76% relative to the F1. Notably, we find that the COTe’s granularity robustness largely holds even without explicit SSU labelling, lowering the barriers to entry for using the system. Finally, we release an SSU labelled dataset and a Python library for applying COTe in DLA projects.

Method: Proposes CognitionCapturerPro with: 1) uncertainty-weighted similarity scoring to quantify modality-specific fidelity, 2) fusion encoder for integrating shared representations from EEG, images, text, depth, and edges, 3) simplified alignment module, and 4) pre-trained diffusion model for generation.

Result: Significantly outperforms original CognitionCapturer on THINGS-EEG dataset, improving Top-1 retrieval accuracy by 25.9% and Top-5 by 10.6%.

Conclusion: The integration of multi-modal priors with uncertainty weighting effectively addresses fidelity loss in EEG-to-visual reconstruction, demonstrating substantial improvements in retrieval accuracy.

Abstract: Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: https://github.com/XiaoZhangYES/CognitionCapturerPro.

Method: Created Dyn-Bench, a large-scale benchmark from diverse real-world and synthetic video datasets with multi-stage filtering. Contains 1k videos, 7k VQA pairs, and 3k dynamic object grounding pairs. Evaluated general, spatial and region-level MLLMs using both linguistic and visual outputs. Proposed structured integration approaches including Mask-Guided Fusion and Spatio-Temporal Textual Cognitive Map (ST-TCM).

Result: Existing models cannot simultaneously maintain strong performance in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent motion interpretations. Conventional prompting strategies provide limited improvement, while structured integration approaches significantly enhance dynamics perception and spatio-temporal reasoning.

Conclusion: Current MLLMs have significant limitations in spatio-temporal reasoning and dynamic object grounding. Structured integration methods like Mask-Guided Fusion and ST-TCM show promise for improving “thinking in dynamics” capabilities, enabling better understanding of physical 4D reality.

Abstract: Humans inhabit a physical 4D world where geometric structure and semantic content evolve over time, constituting a dynamic 4D reality (spatial with temporal dimension). While current Multimodal Large Language Models (MLLMs) excel in static visual understanding, can they also be adept at “thinking in dynamics”, i.e., perceive, track and reason about spatio-temporal dynamics in evolving scenes? To systematically assess their spatio-temporal reasoning and localized dynamics perception capabilities, we introduce Dyn-Bench, a large-scale benchmark built from diverse real-world and synthetic video datasets, enabling robust and scalable evaluation of spatio-temporal understanding. Through multi-stage filtering from massive 2D and 4D data sources, Dyn-Bench provides a high-quality collection of dynamic scenes, comprising 1k videos, 7k visual question answering (VQA) pairs, and 3k dynamic object grounding pairs. We probe general, spatial and region-level MLLMs to express how they think in dynamics both linguistically and visually, and find that existing models cannot simultaneously maintain strong performance in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent interpretations of motion and interaction. Notably, conventional prompting strategies (e.g., chain-of-thought or caption-based hints) provide limited improvement, whereas structured integration approaches, including Mask-Guided Fusion and Spatio-Temporal Textual Cognitive Map (ST-TCM), significantly enhance MLLMs’ dynamics perception and spatio-temporal reasoning in the physical 4D world. Code and benchmark are available at https://dyn-bench.github.io/.

Method: Proposes SLICE framework that decouples image semantics into four factors (subject, environment, action, detail) and precisely anchors them to distinct regions in the initial Gaussian noise of diffusion models. Uses compartmentalized embedding for fine-grained semantic binding.

Result: SLICE significantly outperforms existing baselines against semantic-guided regeneration attacks, reduces attack success while preserving image quality and semantic fidelity. Provides theoretical justification for tamper localization and statistical guarantees on false-accept rates.

Conclusion: SLICE offers a practical, training-free provenance solution that is both fine-grained in diagnosis and robust to realistic adversarial manipulations for diffusion model watermarking.

Abstract: Watermarking the initial noise of diffusion models has emerged as a promising approach for image provenance, but content-independent noise patterns can be forged via inversion and regeneration attacks. Recent semantic-aware watermarking methods improve robustness by conditioning verification on image semantics. However, their reliance on a single global semantic binding makes them vulnerable to localized but globally coherent semantic edits. To address this limitation and provide a trustworthy semantic-aware watermark, we propose $\underline{\textbf{S}}$emantic $\underline{\textbf{L}}$atent $\underline{\textbf{I}}$njection via $\underline{\textbf{C}}$ompartmentalized $\underline{\textbf{E}}$mbedding ($\textbf{SLICE}$). Our framework decouples image semantics into four semantic factors (subject, environment, action, and detail) and precisely anchors them to distinct regions in the initial Gaussian noise. This fine-grained semantic binding enables advanced watermark verification where semantic tampering is detectable and localizable. We theoretically justify why SLICE enables robust and reliable tamper localization and provides statistical guarantees on false-accept rates. Experimental results demonstrate that SLICE significantly outperforms existing baselines against advanced semantic-guided regeneration attacks, substantially reducing attack success while preserving image quality and semantic fidelity. Overall, SLICE offers a practical, training-free provenance solution that is both fine-grained in diagnosis and robust to realistic adversarial manipulations.

Method: Proposes a lightweight post-association correction framework that suppresses unreliable appearance updates under high-overlap conditions using Intersection over Area (IoA)-based filtering, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs.

Result: Achieves 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on MOT17 test set (5.7 FPS), and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on MOT20 test set (0.6 FPS). Produces only 29.55% long-term identity switches, substantially lower than existing online trackers.

Conclusion: The framework effectively mitigates long-term identity switches without global optimization or re-identification, making it suitable for real-time robotic applications while maintaining state-of-the-art performance on MOT benchmarks.

Abstract: Reliable multi-object tracking (MOT) is essential for robotic systems operating in complex and dynamic environments. Despite recent advances in detection and association, online MOT methods remain vulnerable to identity switches caused by frequent occlusions and object overlap, where incorrect associations can propagate over time and degrade tracking reliability. We present a lightweight post-association correction framework (FC-Track) for online MOT that explicitly targets overlap-induced mismatches during inference. The proposed method suppresses unreliable appearance updates under high-overlap conditions using an Intersection over Area (IoA)-based filtering strategy, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs. By preventing short-term mismatches from propagating, our framework effectively mitigates long-term identity switches without resorting to global optimization or re-identification. The framework operates online without global optimization or re-identification, making it suitable for real-time robotic applications. We achieve 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on the MOT17 test set with a running speed of 5.7 FPS, and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on the MOT20 test set with a running speed of 0.6 FPS. Specifically, our framework FC-Track produces only 29.55% long-term identity switches, which is substantially lower than existing online trackers. Meanwhile, our framework maintains state-of-the-art performance on the MOT20 benchmark.

Method: “Show, Don’t Tell” paradigm: trains bespoke object detectors on automatically created datasets supervised by human demonstrations themselves, bypassing language. Integrated on-robot system for automatic dataset creation and novel object detection.

Result: Empirical results show the pipeline significantly outperforms state-of-the-art detection and recognition methods for manipulated objects, leading to improved robot task completion.

Conclusion: The approach enables quick training of customized detectors for relevant objects observed in human demonstrations by bypassing language, eliminating need for prompt engineering.

Abstract: How can a robot quickly identify and recognize new objects shown to it during a human demonstration? Existing closed-set object detectors frequently fail at this because the objects are out-of-distribution. While open-set detectors (e.g., VLMs) sometimes succeed, they often require expensive and tedious human-in-the-loop prompt engineering to uniquely recognize novel object instances. In this paper, we present a self-supervised system that eliminates the need for tedious language descriptions and expensive prompt engineering by training a bespoke object detector on an automatically created dataset, supervised by the human demonstration itself. In our approach, “Show, Don’t Tell,” we show the detector the specific objects of interest during the demonstration, rather than telling the detector about these objects via complex language descriptions. By bypassing language altogether, this paradigm enables us to quickly train bespoke detectors tailored to the relevant objects observed in human task demonstrations. We develop an integrated on-robot system to deploy our “Show, Don’t Tell” paradigm of automatic dataset creation and novel object-detection on a real-world robot. Empirical results demonstrate that our pipeline significantly outperforms state-of-the-art detection and recognition methods for manipulated objects, leading to improved task completion for the robot.

Method: Reformulates panoramic segmentation as fixed-trajectory perspective video segmentation by decomposing panoramas into overlapping perspective patches sampled along continuous spherical traversal. Uses InfiniGen engine to synthesize 183,440 4K-resolution panoramic images with instance segmentation labels for large-scale supervision.

Result: Achieves +17.2 zero-shot mIoU gain over vanilla SAM2 of different sizes on real-world 4K panorama benchmark, demonstrating effective generalization to real-world 360° images.

Conclusion: SAP successfully addresses the degradation of foundation models on 360° panoramas through trajectory-aligned paradigm, enabling high-performance panoramic instance segmentation for embodied and AR applications.

Abstract: Promptable instance segmentation is widely adopted in embodied and AR systems, yet the performance of foundation models trained on perspective imagery often degrades on 360° panoramas. In this paper, we introduce Segment Any 4K Panorama (SAP), a foundation model for 4K high-resolution panoramic instance-level segmentation. We reformulate panoramic segmentation as fixed-trajectory perspective video segmentation, decomposing a panorama into overlapping perspective patches sampled along a continuous spherical traversal. This memory-aligned reformulation preserves native 4K resolution while restoring the smooth viewpoint transitions required for stable cross-view propagation. To enable large-scale supervision, we synthesize 183,440 4K-resolution panoramic images with instance segmentation labels using the InfiniGen engine. Trained under this trajectory-aligned paradigm, SAP generalizes effectively to real-world 360° images, achieving +17.2 zero-shot mIoU gain over vanilla SAM2 of different sizes on real-world 4K panorama benchmark.

Method: Three key components: 1) Unified vision tokenizer encoding image latent states into semantic tokens, 2) LLM-based Transformer unifying autoregressive text decoding and diffusion image decoding, 3) Cascaded flow matching head that decodes visual semantics first then injects semantically gated detail residuals for high-frequency content refinement.

Result: Cheers matches or surpasses advanced unified multimodal models in both visual understanding and generation, achieves 4x token compression enabling efficient high-resolution image encoding/generation, outperforms Tar-1.5B on GenEval and MMBench with only 20% training cost.

Conclusion: Cheers demonstrates effective and efficient unified multimodal modeling by decoupling semantics from details, enabling both high-quality visual understanding and generation within a single model with significant compression benefits.

Abstract: A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.

Method: Proposes High-Fidelity In-Context Learning (HiFICL) with three components: 1) virtual key-value pairs as learnable context, 2) low-rank factorization for stable training, and 3) end-to-end training objective. This creates context-aware Parameter-Efficient Fine-Tuning.

Result: Extensive experiments show HiFICL consistently outperforms existing approximation methods on several multimodal benchmarks.

Conclusion: HiFICL provides a more faithful modeling of in-context learning mechanisms for multimodal models, offering improved performance and stability compared to existing methods.

Abstract: In-Context Learning (ICL) is a significant paradigm for Large Multimodal Models (LMMs), using a few in-context demonstrations (ICDs) for new task adaptation. However, its performance is sensitive to demonstration configurations and computationally expensive. Mathematically, the influence of these demonstrations can be decomposed into a dynamic mixture of the standard attention output and the context values. Current approximation methods simplify this process by learning a “shift vector”. Inspired by the exact decomposition, we introduce High-Fidelity In-Context Learning (HIFICL) to more faithfully model the ICL mechanism. HIFICL consists of three key components: 1) a set of “virtual key-value pairs” to act as a learnable context, 2) a low-rank factorization for stable and regularized training, and 3) a simple end-to-end training objective. From another perspective, this mechanism constitutes a form of context-aware Parameter-Efficient Fine-Tuning (PEFT). Extensive experiments show that HiFICL consistently outperforms existing approximation methods on several multimodal benchmarks. The code is available at https://github.com/bbbandari/HiFICL.

Method: Proposes Hierarchical Dual-Change Collaborative Learning (HDC-CL) with Dynamic Adaptive Layout Transformer (DALT) to model diverse spatial layouts adaptively, learning features from overlapping and non-overlapping regions. Also introduces Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) to enhance sensitivity to viewpoint shift directions.

Result: Achieves state-of-the-art performance on the new UAV-SCC dataset. The method effectively handles viewpoint-induced changes in UAV imagery with partially overlapping content.

Conclusion: The paper introduces a novel UAV scene change captioning task, proposes effective methods for handling viewpoint-induced changes, and provides a benchmark dataset to advance research in this area of dynamic aerial scene understanding.

Abstract: This paper proposes a novel task for UAV scene understanding - UAV Scene Change Captioning (UAV-SCC) - which aims to generate natural language descriptions of semantic changes in dynamic aerial imagery captured from a movable viewpoint. Unlike traditional change captioning that mainly describes differences between image pairs captured from a fixed camera viewpoint over time, UAV scene change captioning focuses on image-pair differences resulting from both temporal and spatial scene variations dynamically captured by a moving camera. The key challenge lies in understanding viewpoint-induced scene changes from UAV image pairs that share only partially overlapping scene content due to viewpoint shifts caused by camera rotation, while effectively exploiting the relative orientation between the two images. To this end, we propose a Hierarchical Dual-Change Collaborative Learning (HDC-CL) method for UAV scene change captioning. In particular, a novel transformer, \emph{i.e.} Dynamic Adaptive Layout Transformer (DALT) is designed to adaptively model diverse spatial layouts of the image pair, where the interrelated features derived from the overlapping and non-overlapping regions are learned within the flexible and unified encoding layer. Furthermore, we propose a Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) method to enhance the model’s sensitivity to viewpoint shift directions, enabling more accurate change captioning. To facilitate in-depth research on this task, we construct a new benchmark dataset, named UAV-SCC dataset, for UAV scene change captioning. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on this task. The dataset and code will be publicly released upon acceptance of this paper.

Method: TerraFlow uses temporal training objectives to enable sequence-aware learning across space, time, and modality. The approach is designed to be robust to variable-length inputs commonly found in real-world Earth observation data.

Result: TerraFlow outperforms state-of-the-art foundation models across all temporal tasks of GEO-Bench-2 benchmark, achieving up to 50% improvement in F1 score and 24% improvement in Brier score. It also shows promising results for natural disaster risk map prediction where other models fail.

Conclusion: TerraFlow represents an effective approach for multimodal, multitemporal learning in Earth observation, demonstrating superior performance on temporal tasks and potential for practical applications like natural disaster risk prediction.

Abstract: We propose TerraFlow, a novel approach to multimodal, multitemporal learning for Earth observation. TerraFlow builds on temporal training objectives that enable sequence-aware learning across space, time, and modality, while remaining robust to the variable-length inputs commonly encountered in real-world Earth observation data. Our experiments demonstrate superiority of TerraFlow over state-of-the-art foundation models for Earth observation across all temporal tasks of the GEO-Bench-2 benchmark. We additionally demonstrate that TerraFlow is able to make initial steps towards deep-learning based risk map prediction for natural disasters – a task on which other state-of-the-art foundation models frequently collapse. TerraFlow outperforms state-of-the-art foundation models by up to 50% in F1 score and 24% in Brier score.

Method: Integrates four modalities: scene (VideoMAE), face (emotional embeddings with statistical pooling), audio (EmotionWav2Vec2.0 with Mamba encoder), and text (fine-tuned transformers). Uses multimodal fusion with prototype-augmented variants.

Result: Multimodal fusion outperforms unimodal baselines: best unimodal MF1 70.02%, best multimodal 83.25%. Ensemble achieved 71.43% on final test.

Conclusion: Complementary multimodal cues and robust fusion strategies are crucial for effective ambivalence/hesitancy recognition in videos.

Abstract: Ambivalence/hesitancy recognition in unconstrained videos is a challenging problem due to the subtle, multimodal, and context-dependent nature of this behavioral state. In this paper, a multimodal approach for video-level ambivalence/hesitancy recognition is presented for the 10th ABAW Competition. The proposed approach integrates four complementary modalities: scene, face, audio, and text. Scene dynamics are captured with a VideoMAE-based model, facial information is encoded through emotional frame-level embeddings aggregated by statistical pooling, acoustic representations are extracted with EmotionWav2Vec2.0 and processed by a Mamba-based temporal encoder, and linguistic cues are modeled using fine-tuned transformer-based text models. The resulting unimodal embeddings are further combined using multimodal fusion models, including prototype-augmented variants. Experiments on the BAH corpus demonstrate clear gains of multimodal fusion over all unimodal baselines. The best unimodal configuration achieved an average MF1 of 70.02%, whereas the best multimodal fusion model reached 83.25%. The highest final test performance, 71.43%, was obtained by an ensemble of five prototype-augmented fusion models. The obtained results highlight the importance of complementary multimodal cues and robust fusion strategies for ambivalence/hesitancy recognition.

Method: Proposes SAVA-X with an Align-Fuse-Detect framework featuring: (1) view-conditioned adaptive sampling, (2) scene-adaptive view embeddings, and (3) bidirectional cross-attention fusion to handle cross-view video analysis.

Result: On the EgoMe benchmark, SAVA-X consistently improves AUPRC and mean tIoU over all baselines adapted from dense video captioning and temporal action detection, with ablations confirming the complementary benefits of its components.

Conclusion: SAVA-X effectively addresses the challenging Ego→Exo imitation error detection problem by handling cross-view domain shift and temporal misalignment through its novel framework components.

Abstract: Error detection is crucial in industrial training, healthcare, and assembly quality control. Most existing work assumes a single-view setting and cannot handle the practical case where a third-person (exo) demonstration is used to assess a first-person (ego) imitation. We formalize Ego$\rightarrow$Exo Imitation Error Detection: given asynchronous, length-mismatched ego and exo videos, the model must localize procedural steps on the ego timeline and decide whether each is erroneous. This setting introduces cross-view domain shift, temporal misalignment, and heavy redundancy. Under a unified protocol, we adapt strong baselines from dense video captioning and temporal action detection and show that they struggle in this cross-view regime. We then propose SAVA-X, an Align-Fuse-Detect framework with (i) view-conditioned adaptive sampling, (ii) scene-adaptive view embeddings, and (iii) bidirectional cross-attention fusion. On the EgoMe benchmark, SAVA-X consistently improves AUPRC and mean tIoU over all baselines, and ablations confirm the complementary benefits of its components. Code is available at https://github.com/jack1ee/SAVAX.

Method: Proposes online RL variant that samples paired trajectories and pulls flow velocity toward more favorable images. Treats entire sampling process as single action rather than separate steps. Uses high-quality vision language models and off-the-shelf quality metrics for rewards.

Result: Method converges faster than previous approaches and yields higher output quality and prompt alignment. Evaluated using broad set of metrics showing improved performance.

Conclusion: The proposed online RL approach with paired trajectories and treating sampling as single action provides more stable training and better results for diffusion model post-training.

Abstract: Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.

Method: Two key components: 1) Anchor-based Motion Guidance (AMG) builds structurally stable anchors from original and edited Gaussians, using optimal transport for correspondence to enable consistent deformation propagation. 2) Color Uncertainty-guided Appearance Refinement (CUAR) preserves temporal appearance consistency by estimating per-Gaussian color uncertainty and selectively refining occlusion-prone regions.

Result: Extensive experiments show Catalyst4D achieves temporally stable, high-fidelity dynamic scene editing and outperforms existing methods in both visual quality and motion coherence.

Conclusion: Catalyst4D provides an effective framework for transferring high-quality 3D edits to dynamic 4D Gaussian scenes while maintaining spatial and temporal coherence, addressing key challenges in dynamic scene editing.

Abstract: Recent advances in 3D scene editing using NeRF and 3DGS enable high-quality static scene editing. In contrast, dynamic scene editing remains challenging, as methods that directly extend 2D diffusion models to 4D often produce motion artifacts, temporal flickering, and inconsistent style propagation. We introduce Catalyst4D, a framework that transfers high-quality 3D edits to dynamic 4D Gaussian scenes while maintaining spatial and temporal coherence. At its core, Anchor-based Motion Guidance (AMG) builds a set of structurally stable and spatially representative anchors from both original and edited Gaussians. These anchors serve as robust region-level references, and their correspondences are established via optimal transport to enable consistent deformation propagation without cross-region interference or motion drift. Complementarily, Color Uncertainty-guided Appearance Refinement (CUAR) preserves temporal appearance consistency by estimating per-Gaussian color uncertainty and selectively refining regions prone to occlusion-induced artifacts. Extensive experiments demonstrate that Catalyst4D achieves temporally stable, high-fidelity dynamic scene editing and outperforms existing methods in both visual quality and motion coherence.

Method: Plug-in Visual Injection (PVI) attaches to pretrained action experts via zero-initialized residual pathways, injecting auxiliary visual representations (temporal video features like V-JEPA2) without major architectural changes.

Result: PVI achieves consistent gains over base policies and alternative injection strategies, with temporal video features outperforming static image features, especially on multi-phase tasks requiring state tracking.

Conclusion: PVI provides a practical, lightweight approach to enhance VLA manipulation policies with temporal visual information, demonstrating effectiveness in both simulation and real-robot bimanual cloth folding tasks.

Abstract: VLA architectures that pair a pretrained VLM with a flow-matching action expert have emerged as a strong paradigm for language-conditioned manipulation. Yet the VLM, optimized for semantic abstraction and typically conditioned on static visual observations, tends to attenuate fine-grained geometric cues and often lacks explicit temporal evidence for the action expert. Prior work mitigates this by injecting auxiliary visual features, but existing approaches either focus on static spatial representations or require substantial architectural modifications to accommodate temporal inputs, leaving temporal information underexplored. We propose Plug-in Visual Injection (PVI), a lightweight, encoder-agnostic module that attaches to a pretrained action expert and injects auxiliary visual representations via zero-initialized residual pathways, preserving pretrained behavior with only single-stage fine-tuning. Using PVI, we obtain consistent gains over the base policy and a range of competitive alternative injection strategies, and our controlled study shows that temporal video features (V-JEPA2) outperform strong static image features (DINOv2), with the largest gains on multi-phase tasks requiring state tracking and coordination. Real-robot experiments on long-horizon bimanual cloth folding further demonstrate the practicality of PVI beyond simulation.

Method: Proposes FedBPrompt with Body Distribution Aware Visual Prompts Mechanism (BAPM) using holistic full body prompts and body part alignment prompts. Also includes Prompt-based Fine-Tuning Strategy (PFTS) that freezes ViT backbone and updates only lightweight prompts.

Result: BAPM enhances feature discrimination and cross-domain generalization. PFTS achieves notable performance gains within few aggregation rounds while significantly reducing communication overhead.

Conclusion: FedBPrompt is a flexible and effective solution for federated person re-identification that can be easily integrated into existing ViT-based FedDG-ReID frameworks.

Abstract: Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints – a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at https://github.com/leavlong/FedBPrompt.

Method: Created the largest BSA dataset with 10 basic actions across 6 surgical specialties (11,000+ video clips). Developed a foundation model for general-purpose BSA recognition, validated cross-specialty performance, and integrated with large vision-language models for downstream applications like skill assessment and action planning.

Result: The model demonstrated robust cross-specialty performance across different procedural types and body parts. Downstream applications showed successful surgical skill assessment in prostatectomy and action planning in cholecystectomy/nephrectomy. Multinational surgeons validated the clinical relevance of the language model’s action planning explanations.

Conclusion: Basic surgical actions can be robustly recognized across scenarios, and accurate BSA understanding models can facilitate complex surgical applications and accelerate the development of surgical superintelligence.

Abstract: Artificial intelligence, imaging, and large language models have the potential to transform surgical practice, training, and automation. Understanding and modeling of basic surgical actions (BSA), the fundamental unit of operation in any surgery, is important to drive the evolution of this field. In this paper, we present a BSA dataset comprising 10 basic actions across 6 surgical specialties with over 11,000 video clips, which is the largest to date. Based on the BSA dataset, we developed a new foundation model that conducts general-purpose recognition of basic actions. Our approach demonstrates robust cross-specialist performance in experiments validated on datasets from different procedural types and various body parts. Furthermore, we demonstrate downstream applications enabled by the BAS foundation model through surgical skill assessment in prostatectomy using domain-specific knowledge, and action planning in cholecystectomy and nephrectomy using large vision-language models. Multinational surgeons’ evaluation of the language model’s output of the action planning explainable texts demonstrated clinical relevance. These findings indicate that basic surgical actions can be robustly recognized across scenarios, and an accurate BSA understanding model can essentially facilitate complex applications and speed up the realization of surgical superintelligence.

Method: Proposes a structure-faithful framework using semantic anchors (stakes) derived from language-driven attribute descriptions encoded by semantic encoders like CLIP. Uses structure-aware alignment to maintain knowledge organization before/after unlearning, and structure-aware regularization to regulate updates to critical parameters.

Result: Shows average gains of 32.9% in image classification, 22.5% in retrieval, and 19.3% in face recognition performance, effectively balancing the deletion-retention trade-off and enhancing generalization.

Conclusion: The proposed structure-faithful framework successfully addresses structural collapse in machine unlearning by preserving semantic relations through semantic anchors, achieving better deletion-retention balance and improved generalization across vision tasks.

Abstract: Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion-retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion-retention trade-off and enhancing generalization.

Method: Proposes an Entity-Aware Reasoning (EAR) framework built on visual-linguistic foundation models that generates structured reasoning traces and subject-object grounding outputs. Uses supervised fine-tuning for initialization and entity-aware reward-driven Group Relative Policy Optimization (GRPO) for further optimization.

Result: Extensive experiments on the new ME-RSRG benchmark demonstrate the challenges of multi-entity reasoning and verify the effectiveness of the proposed EAR framework.

Conclusion: The work introduces a new benchmark for multi-entity reasoning in remote sensing and shows that explicit reasoning frameworks can effectively address complex grounding tasks in this domain.

Abstract: Recent advances in reasoning language models and reinforcement learning with verifiable rewards have significantly enhanced multi-step reasoning capabilities. This progress motivates the extension of reasoning paradigms to remote sensing visual grounding task. However, existing remote sensing grounding methods remain largely confined to perception-level matching and single-entity formulations, limiting the role of explicit reasoning and inter-entity modeling. To address this challenge, we introduce a new benchmark dataset for Multi-Entity Reasoning Grounding in Remote Sensing (ME-RSRG). Based on ME-RSRG, we reformulate remote sensing grounding as a multi-entity reasoning task and propose an Entity-Aware Reasoning (EAR) framework built upon visual-linguistic foundation models. EAR generates structured reasoning traces and subject-object grounding outputs. It adopts supervised fine-tuning for cold-start initialization and is further optimized via entity-aware reward-driven Group Relative Policy Optimization (GRPO). Extensive experiments on ME-RSRG demonstrate the challenges of multi-entity reasoning and verify the effectiveness of our proposed EAR framework. Our dataset, code, and models will be available at https://github.com/CV-ShuchangLyu/ME-RSRG.

Method: Watch-Think-Speak paradigm with incremental reasoning updates; Reasoning-Compressed Streaming Memory (RCSM) that compresses reasoning traces into semantic memory; Streaming Reinforcement Learning with Verifiable Rewards to align reasoning timing.

Result: Outperforms existing online video models on multiple streaming video benchmarks while maintaining low latency and memory usage.

Conclusion: ThinkStream provides an effective framework for real-time video understanding in streaming scenarios, enabling interactive multimodal agents to operate in dynamic environments.

Abstract: Real-time understanding of continuous video streams is essential for interactive assistants and multimodal agents operating in dynamic environments. However, most existing video reasoning approaches follow a batch paradigm that defers reasoning until the full video context is observed, resulting in high latency and growing computational cost that are incompatible with streaming scenarios. In this paper, we introduce ThinkStream, a framework for streaming video reasoning based on a Watch–Think–Speak paradigm that enables models to incrementally update their understanding as new video observations arrive. At each step, the model performs a short reasoning update and decides whether sufficient evidence has accumulated to produce a response. To support long-horizon streaming, we propose Reasoning-Compressed Streaming Memory (RCSM), which treats intermediate reasoning traces as compact semantic memory that replaces outdated visual tokens while preserving essential context. We further train the model using a Streaming Reinforcement Learning with Verifiable Rewards scheme that aligns incremental reasoning and response timing with the requirements of streaming interaction. Experiments on multiple streaming video benchmarks show that ThinkStream significantly outperforms existing online video models while maintaining low latency and memory usage. Code, models and data will be released at https://github.com/johncaged/ThinkStream

Method: Integrates geometric and human priors from Pi3X and Multi-HMR into a single trainable neural network, adds scale adjustment module to resolve human-scene scale discrepancy, uses multi-view fusion strategy to aggregate per-view estimates, and employs geometry-based multi-person association method.

Result: Achieves competitive performance in global human motion and multi-view pose estimation on EMDB, RICH, EgoHumans, and EgoExo4D datasets while running over 8x faster than prior optimization-based multi-view approaches.

Conclusion: CHROMM provides an efficient unified framework for joint camera, scene, and human mesh estimation from multi-person multi-view videos without external dependencies.

Abstract: Recent advances in 3D foundation models have led to growing interest in reconstructing humans and their surrounding environments. However, most existing approaches focus on monocular inputs, and extending them to multi-view settings requires additional overhead modules or preprocessed data. To this end, we present CHROMM, a unified framework that jointly estimates cameras, scene point clouds, and human meshes from multi-person multi-view videos without relying on external modules or preprocessing. We integrate strong geometric and human priors from Pi3X and Multi-HMR into a single trainable neural network architecture, and introduce a scale adjustment module to solve the scale discrepancy between humans and the scene. We also introduce a multi-view fusion strategy to aggregate per-view estimates into a single representation at test-time. Finally, we propose a geometry-based multi-person association method, which is more robust than appearance-based approaches. Experiments on EMDB, RICH, EgoHumans, and EgoExo4D show that CHROMM achieves competitive performance in global human motion and multi-view pose estimation while running over 8x faster than prior optimization-based multi-view approaches. Project page: https://nstar1125.github.io/chromm.

Method: Proposes Spectral Defense with two components: 1) 3D frequency filter to selectively prune Gaussians exhibiting abnormally high frequencies, and 2) 2D spectral regularization on renderings that distinguishes naturally isotropic frequencies while penalizing anisotropic angular energy to constrain noisy patterns.

Result: Defense suppresses Gaussian overgrowth by up to 5.92×, reduces memory usage by up to 3.66×, and improves rendering speed by up to 4.34× under attacks while maintaining robust and accurate 3DGS performance.

Conclusion: Spectral Defense effectively protects 3DGS from resource-targeting attacks by addressing the spectral vulnerabilities that spatial-domain methods miss, providing comprehensive security improvements in growth suppression, memory reduction, and speed enhancement.

Abstract: Recent advances in 3D Gaussian Splatting (3DGS) deliver high-quality rendering, yet the Gaussian representation exposes a new attack surface, the resource-targeting attack. This attack poisons training images, excessively inducing Gaussian growth to cause resource exhaustion. Although efficiency-oriented methods such as smoothing, thresholding, and pruning have been explored, these spatial-domain strategies operate on visible structures but overlook how stealthy perturbations distort the underlying spectral behaviors of training data. As a result, poisoned inputs introduce abnormal high-frequency amplifications that mislead 3DGS into interpreting noisy patterns as detailed structures, ultimately causing unstable Gaussian overgrowth and degraded scene fidelity. To address this, we propose \textbf{Spectral Defense} in Gaussian and image fields. We first design a 3D frequency filter to selectively prune Gaussians exhibiting abnormally high frequencies. Since natural scenes also contain legitimate high-frequency structures, directly suppressing high frequencies is insufficient, and we further develop a 2D spectral regularization on renderings, distinguishing naturally isotropic frequencies while penalizing anisotropic angular energy to constrain noisy patterns. Experiments show that our defense builds robust, accurate, and secure 3DGS, suppressing overgrowth by up to $5.92\times$, reducing memory by up to $3.66\times$, and improving speed by up to $4.34\times$ under attacks.

Method: The authors first analyze adversarially fine-tuned models to discover that robustness mechanisms are localized in shallow layers (driven by low-frequency spectral bias and input-insensitive attention patterns). Based on this insight, they propose R-Adapt, which freezes all pre-trained weights and introduces minimal adaptations only in the initial layers, supporting training-free, model-guided, and data-driven paradigms.

Result: Extensive evaluations on 18 datasets and diverse tasks demonstrate state-of-the-art performance under various attacks. R-Adapt generalizes efficiently to large VLMs like LLaVA and Qwen-VL to enhance their robustness while maintaining clean accuracy.

Conclusion: The paper shows that adversarial robustness in VLMs can be achieved with minimal architectural changes by focusing adaptations on shallow layers, providing a flexible framework that balances robustness and accuracy better than existing methods.

Abstract: Achieving adversarial robustness in Vision-Language Models (VLMs) inevitably compromises accuracy on clean data, presenting a long-standing and challenging trade-off. In this work, we revisit this trade-off by investigating a fundamental question: What makes VLMs robust? Through a detailed analysis of adversarially fine-tuned models, we examine how robustness mechanisms function internally and how they interact with clean accuracy. Our analysis reveals that adversarial robustness is not uniformly distributed across network depth. Instead, unexpectedly, it is primarily localized within the shallow layers, driven by a low-frequency spectral bias and input-insensitive attention patterns. Meanwhile, updates to the deep layers tend to undermine both clean accuracy and robust generalization. Motivated by these insights, we propose Adversarial Robustness Adaptation (R-Adapt), a simple yet effective framework that freezes all pre-trained weights and introduces minimal, insight-driven adaptations only in the initial layers. This design achieves an exceptional balance between adversarial robustness and clean accuracy. R-Adapt further supports training-free, model-guided, and data-driven paradigms, offering flexible pathways to seamlessly equip standard models with robustness. Extensive evaluations on 18 datasets and diverse tasks demonstrate our state-of-the-art performance under various attacks. Notably, R-Adapt generalizes efficiently to large vision-language models (e.g., LLaVA and Qwen-VL) to enhance their robustness. Our project page is available at https://summu77.github.io/R-Adapt.

Method: Attention-based Multiple Instance Learning (MIL) on ConvNeXt backbone with Gradient Reversal Layer (GRL) for adversarial gender debiasing, focal loss with label smoothing, stratified cross-validation, targeted oversampling, and ensemble inference with test-time augmentation

Result: Achieved mean validation competition score of 0.685 (std 0.030) with best single fold reaching 0.759 on the Fair Disease Diagnosis Challenge benchmark

Conclusion: Proposed framework effectively addresses fairness concerns in medical imaging while maintaining diagnostic performance, with publicly available code for reproducibility

Abstract: We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories – Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma – with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at https://github.com/ADE-17/cvpr-fair-chest-ct

Method: Proposes OARS, a process-aware online alignment framework built on COMPASS, a MLLM-based reward that evaluates LR to SR transitions by jointly modeling fidelity preservation and perceptual gain with input-quality-adaptive trade-off. Uses three-stage perceptual annotation pipeline to curate COMPASS-20K dataset. Performs progressive online alignment from cold-start flow matching to full-reference and finally reference-free RL via shallow LoRA optimization for on-policy exploration.

Result: Extensive experiments and user studies demonstrate consistent perceptual improvements while maintaining fidelity, achieving state-of-the-art performance on Real-ISR benchmarks.

Conclusion: OARS provides an effective online alignment framework for real-world image super-resolution that successfully balances perceptual quality and fidelity through MLLM-based reward guidance and progressive reinforcement learning.

Abstract: Aligning generative real-world image super-resolution models with human visual preference is challenging due to the perception–fidelity trade-off and diverse, unknown degradations. Prior approaches rely on offline preference optimization and static metric aggregation, which are often non-interpretable and prone to pseudo-diversity under strong conditioning. We propose OARS, a process-aware online alignment framework built on COMPASS, a MLLM-based reward that evaluates the LR to SR transition by jointly modeling fidelity preservation and perceptual gain with an input-quality-adaptive trade-off. To train COMPASS, we curate COMPASS-20K spanning synthetic and real degradations, and introduce a three-stage perceptual annotation pipeline that yields calibrated, fine-grained training labels. Guided by COMPASS, OARS performs progressive online alignment from cold-start flow matching to full-reference and finally reference-free RL via shallow LoRA optimization for on-policy exploration. Extensive experiments and user studies demonstrate consistent perceptual improvements while maintaining fidelity, achieving state-of-the-art performance on Real-ISR benchmarks.

Method: Four specialized agents collaborate: Interpreter parses prompts into attribute-rich object descriptors and decides generation pathway; Planner uses divide-and-conquer strategy for incremental layout planning; Checker validates spatial consistency and attribute alignment; Painter synthesizes images step-by-step with visual context.

Result: Extensive experiments on GenEval and DPG-Bench benchmarks show coDrawAgents substantially improves text-image alignment, spatial accuracy, and attribute binding compared to existing methods.

Conclusion: The multi-agent dialogue framework effectively addresses key challenges in compositional text-to-image generation through collaborative agents that reduce layout complexity, ground planning in visual context, and enable explicit error correction.

Abstract: Text-to-image generation has advanced rapidly, but existing models still struggle with faithfully composing multiple objects and preserving their attributes in complex scenes. We propose coDrawAgents, an interactive multi-agent dialogue framework with four specialized agents: Interpreter, Planner, Checker, and Painter that collaborate to improve compositional generation. The Interpreter adaptively decides between a direct text-to-image pathway and a layout-aware multi-agent process. In the layout-aware mode, it parses the prompt into attribute-rich object descriptors, ranks them by semantic salience, and groups objects with the same semantic priority level for joint generation. Guided by the Interpreter, the Planner adopts a divide-and-conquer strategy, incrementally proposing layouts for objects with the same semantic priority level while grounding decisions in the evolving visual context of the canvas. The Checker introduces an explicit error-correction mechanism by validating spatial consistency and attribute alignment, and refining layouts before they are rendered. Finally, the Painter synthesizes the image step by step, incorporating newly planned objects into the canvas to provide richer context for subsequent iterations. Together, these agents address three key challenges: reducing layout complexity, grounding planning in visual context, and enabling explicit error correction. Extensive experiments on benchmarks GenEval and DPG-Bench demonstrate that coDrawAgents substantially improves text-image alignment, spatial accuracy, and attribute binding compared to existing methods.

Method: Created SortScrews dataset with 560 RGB images at 512×512 resolution covering 6 screw types and background class. Used standardized acquisition setup with mild lighting and perspective variations. Provided reusable data collection script. Established baselines using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet.

Result: Lightweight models achieved strong classification accuracy despite limited dataset size, demonstrating that controlled acquisition conditions enable effective learning with relatively small datasets.

Conclusion: SortScrews dataset fills gap in publicly available screw classification datasets, provides reproducible research framework, and shows effective visual classification possible with controlled acquisition and transfer learning.

Abstract: Automatic identification of screw types is important for industrial automation, robotics, and inventory management. However, publicly available datasets for screw classification are scarce, particularly for controlled single-object scenarios commonly encountered in automated sorting systems. In this work, we introduce $\textbf{SortScrews}$, a dataset for casewise visual classification of screws. The dataset contains 560 RGB images at $512\times512$ resolution covering six screw types and a background class. Images are captured using a standardized acquisition setup and include mild variations in lighting and camera perspective across four capture settings. To facilitate reproducible research and dataset expansion, we also provide a reusable data collection script that allows users to easily construct similar datasets for custom hardware components using inexpensive camera setups. We establish baseline results using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet. In addition, we conduct a well-explored failure analysis. Despite the limited dataset size, these lightweight models achieve strong classification accuracy, demonstrating that controlled acquisition conditions enable effective learning even with relatively small datasets. The dataset, collection pipeline, and baseline training code are publicly available at https://github.com/ATATC/SortScrews.

Method: ERBA uses two-stage conditioning: Molecular Recognition Cross-Attention (MRCA) for substrate information injection, and Geometry-aware Mixture-of-Experts (G-MoE) for active-site structure integration. Enzyme-Substrate Distribution Alignment (ESDA) maintains semantic fidelity in PLM manifold.

Result: ERBA achieves consistent gains across three kinetic endpoints and multiple PLM backbones, with stronger out-of-distribution performance compared to sequence-only and shallow-fusion baselines.

Conclusion: ERBA provides a biologically grounded approach to scalable kinetic prediction and offers a foundation for incorporating additional factors like cofactors, mutations, and time-resolved structural cues.

Abstract: Predicting enzyme kinetic parameters quantifies how efficiently an enzyme catalyzes a specific substrate under defined biochemical conditions. Canonical parameters such as the turnover number ($k_\text{cat}$), Michaelis constant ($K_\text{m}$), and inhibition constant ($K_\text{i}$) depend jointly on the enzyme sequence, the substrate chemistry, and the conformational adaptation of the active site during binding. Many learning pipelines simplify this process to a static compatibility problem between the enzyme and substrate, fusing their representations through shallow operations and regressing a single value. Such formulations overlook the staged nature of catalysis, which involves both substrate recognition and conformational adaptation. In this regard, we reformulate kinetic prediction as a staged multimodal conditional modeling problem and introduce the Enzyme-Reaction Bridging Adapter (ERBA), which injects cross-modal information via fine-tuning into Protein Language Models (PLMs) while preserving their biochemical priors. ERBA performs conditioning in two stages: Molecular Recognition Cross-Attention (MRCA) first injects substrate information into the enzyme representation to capture specificity; Geometry-aware Mixture-of-Experts (G-MoE) then integrates active-site structure and routes samples to pocket-specialized experts to reflect induced fit. To maintain semantic fidelity, Enzyme-Substrate Distribution Alignment (ESDA) enforces distributional consistency within the PLM manifold in a reproducing kernel Hilbert space. Experiments across three kinetic endpoints and multiple PLM backbones, ERBA delivers consistent gains and stronger out-of-distribution performance compared with sequence-only and shallow-fusion baselines, offering a biologically grounded route to scalable kinetic prediction and a foundation for adding cofactors, mutations, and time-resolved structural cues.

Method: Proposes a hierarchical classification framework with three levels: (1) state detection (new vs. worn), (2) wear type identification (rectangular, concave, convex) and flap tear detection, (3) severity assessment (partial vs. complete deformation). Uses custom-built dataset of real flap wheel images and transfer learning with EfficientNetV2 architecture. Employs Grad-CAM for model interpretability.

Result: High robustness with classification accuracies ranging from 93.8% (flap tears) to 99.3% (concave severity). Grad-CAM validation shows models learn physically relevant features and helps examine false classifications.

Conclusion: The hierarchical method provides basis for adaptive process control and wear consideration in automated flap wheel grinding, enabling automated monitoring of complex wear patterns in industrial applications.

Abstract: Abrasive flap wheels are common for finishing complex free-form surfaces due to their flexibility. However, this flexibility results in complex wear patterns such as concave/convex flap profiles or flap tears, which influence the grinding result. This paper proposes a novel, vision-based hierarchical classification framework to automate the wear condition monitoring of flap wheels. Unlike monolithic classification approaches, we decompose the problem into three logical levels: (1) state detection (new vs. worn), (2) wear type identification (rectangular, concave, convex) and flap tear detection, and (3) severity assessment (partial vs. complete deformation). A custom-built dataset of real flap wheel images was generated and a transfer learning approach with EfficientNetV2 architecture was used. The results demonstrate high robustness with classification accuracies ranging from 93.8% (flap tears) to 99.3% (concave severity). Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) is utilized to validate that the models learn physically relevant features and examine false classifications. The proposed hierarchical method provides a basis for adaptive process control and wear consideration in automated flap wheel grinding.

Method: Proposes noise-level identity injection for pose-agnostic identity generation from single reference images, and hierarchical dual-branch action control mechanism for improved action controllability, enabling disentangled control over traffic factors.

Result: Superior controllable generation quality with 17% FVD improvement for identity editing, 30% reduction in rotation errors, 47% reduction in translation errors for action control. Downstream stress-testing shows 173% average increase in collision rates.

Conclusion: CompoSIA enables systematic synthesis of adversarial driving scenarios by disentangling traffic factors, revealing substantial planner failures and demonstrating the importance of compositional control for safety-critical edge case generation.

Abstract: A major challenge in autonomous driving is the “long tail” of safety-critical edge cases, which often emerge from unusual combinations of common traffic elements. Synthesizing these scenarios is crucial, yet current controllable generative models provide incomplete or entangled guidance, preventing the independent manipulation of scene structure, object identity, and ego actions. We introduce CompoSIA, a compositional driving video simulator that disentangles these traffic factors, enabling fine-grained control over diverse adversarial driving scenarios. To support controllable identity replacement of scene elements, we propose a noise-level identity injection, allowing pose-agnostic identity generation across diverse element poses, all from a single reference image. Furthermore, a hierarchical dual-branch action control mechanism is introduced to improve action controllability. Such disentangled control enables adversarial scenario synthesis-systematically combining safe elements into dangerous configurations that entangled generators cannot produce. Extensive comparisons demonstrate superior controllable generation quality over state-of-the-art baselines, with a 17% improvement in FVD for identity editing and reductions of 30% and 47% in rotation and translation errors for action control. Furthermore, downstream stress-testing reveals substantial planner failures: across editing modalities, the average collision rate of 3s increases by 173%.

Method: Conducted controlled empirical study comparing 11 SMAs and GP-VMs using unified training/evaluation protocol across three heterogeneous medical imaging datasets covering different modalities, class structures, and data characteristics. Analyzed both segmentation accuracy and qualitative Grad-CAM visualizations for explainability behavior.

Result: GP-VMs outperformed majority of specialized MIS models across all analyzed datasets. Explainability analyses showed GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design.

Conclusion: GP-VMs represent a viable alternative to domain-specific methods for medical image segmentation, highlighting the importance of informed model selection rather than assuming specialized architectures are always superior.

Abstract: Medical image segmentation (MIS) is a fundamental component of computer-assisted diagnosis and clinical decision support systems. Over the past decade, numerous architectures specifically tailored to medical imaging have emerged to address domain-specific challenges such as low contrast, small anatomical structures, and limited annotated data. In parallel, rapid progress in computer vision has produced highly capable general-purpose vision models (GP-VMs) originally designed for natural images. Despite their strong performance on standard vision benchmarks, their effectiveness for MIS remains insufficiently understood. In this work, we conduct a controlled empirical study to examine whether specialized medical segmentation architectures (SMAs) provide systematic advantages over modern GP-VMs for 2D MIS. We compare eleven SMAs and GP-VMs using a unified training and evaluation protocol. Experiments are performed across three heterogeneous datasets covering different imaging modalities, class structures, and data characteristics. Beyond segmentation accuracy, we analyze qualitative Grad-CAM visualizations to investigate explainability (XAI) behavior. Our results demonstrate that, for the analyzed datasets, GP-VMs out-perform the majority of specialized MIS models. Moreover, XAI analyses indicate that GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design. These findings suggest that GP-VMs can represent a viable alternative to domain-specific methods, highlighting the importance of informed model selection for end-to-end MIS systems. All code and resources are available at GitHub.

Method: Three-component framework: (1) adaptive diffusion initialization using specialized algorithms (MPE, TPE, MDM) to identify handle points, target points, and editing regions; (2) guided diffusion encoding for precise point movement; (3) masked region replacement with specialized loss to minimize feature alterations after diffusion.

Result: Superior performance over state-of-the-art methods with >5 dB improvement in PSNR and 5% higher extraction accuracy after cross-media transmission. Broad generalizability across multiple languages and fonts.

Conclusion: TRACE demonstrates effective structure-aware character encoding using diffusion models, making it suitable for practical document security applications with robust noise resistance and cross-language adaptability.

Abstract: We propose TRACE, a structure-aware framework leveraging diffusion models for localized character encoding to embed data. Unlike existing methods that rely on edge features or pre-defined codebooks, TRACE exploits character structures that provide inherent resistance to noise interference due to their stability and unified representation across diverse characters. Our framework comprises three key components: (1) adaptive diffusion initialization that automatically identifies handle points, target points, and editing regions through specialized algorithms including movement probability estimator (MPE), target point estimation (TPE) and mask drawing model (MDM), (2) guided diffusion encoding for precise movement of selected point, and (3) masked region replacement with a specialized loss function to minimize feature alterations after the diffusion process. Comprehensive experiments demonstrate \name{}’s superior performance over state-of-the-art methods, achieving more than 5 dB improvement in PSNR and 5% higher extraction accuracy following cross-media transmission. \name{} achieves broad generalizability across multiple languages and fonts, making it particularly suitable for practical document security applications.

Method: Three-modality approach: 1) Face modality uses GRADA-based embeddings with Transformer temporal regression; 2) Behavior modality uses Qwen3-VL-4B-Instruct for video segment analysis with Mamba for temporal modeling; 3) Audio modality uses WavLM-Large with attention-statistics pooling and cross-modal filtering. Two fusion strategies explored: Directed Cross-Modal Mixture-of-Experts and Reliability-Aware Audio-Visual Fusion.

Result: Achieves Concordance Correlation Coefficient (CCC) of 0.658 on Aff-Wild2 development set, demonstrating state-of-the-art performance for in-the-wild emotion recognition.

Conclusion: The proposed multimodal fusion strategy effectively combines complementary information from face, behavior, and audio modalities, showing strong performance for continuous emotion recognition under challenging in-the-wild conditions.

Abstract: Continuous emotion recognition in terms of valence and arousal under in-the-wild (ITW) conditions remains a challenging problem due to large variations in appearance, head pose, illumination, occlusions, and subject-specific patterns of affective expression. We present a multimodal method for valence-arousal estimation ITW. Our method combines three complementary modalities: face, behavior, and audio. The face modality relies on GRADA-based frame-level embeddings and Transformer-based temporal regression. We use Qwen3-VL-4B-Instruct to extract behavior-relevant information from video segments, while Mamba is used to model temporal dynamics across segments. The audio modality relies on WavLM-Large with attention-statistics pooling and includes a cross-modal filtering stage to reduce the influence of unreliable or non-speech segments. To fuse modalities, we explore two fusion strategies: a Directed Cross-Modal Mixture-of-Experts Fusion Strategy that learns interactions between modalities with adaptive weighting, and a Reliability-Aware Audio-Visual Fusion Strategy that combines visual features at the frame-level while using audio as complementary context. The results are reported on the Aff-Wild2 dataset following the 10th Affective Behavior Analysis in-the-Wild (ABAW) challenge protocol. Experiments demonstrate that the proposed multimodal fusion strategy achieves a Concordance Correlation Coefficient (CCC) of 0.658 on the Aff-Wild2 development set.

Method: Three-step protocol: 1) Select reference staining conditions, 2) Characterize test set staining properties, 3) Apply CPath models under simulated reference staining conditions. Applied to 306 MSI classification models including attention-based multiple instance learning models with various feature extractors.

Result: Classification performance ranged from AUC 0.769-0.911 across models and staining conditions. Robustness ranged from 0.007-0.079, showing weak inverse correlation with classification performance (Pearson r=-0.22).

Conclusion: The evaluation protocol enables robustness-informed CPath model selection and provides insight into performance shifts across H&E staining conditions, supporting identification of operational ranges for reliable model deployment.

Abstract: Sensitivity to staining variation remains a major barrier to deploying computational pathology (CPath) models as hematoxylin and eosin (H&E) staining varies across laboratories, requiring systematic assessment of how this variability affects model prediction. In this work, we developed a three-step protocol for evaluating robustness to H&E staining variation in CPath models. Step 1: Select reference staining conditions, Step 2: Characterize test set staining properties, Step 3: Apply CPath model(s) under simulated reference staining conditions. Here, we first created a new reference staining library based on the PLISM dataset. As an exemplary use case, we applied the protocol to assess the robustness properties of 306 microsatellite instability (MSI) classification models on the unseen SurGen colorectal cancer dataset (n=738), including 300 attention-based multiple instance learning models trained on the TCGA-COAD/READ datasets across three feature extractors (UNI2-h, H-Optimus-1, Virchow2), alongside six public MSI classification models. Classification performance was measured as AUC, and robustness as the min-max AUC range across four simulated staining conditions (low/high H&E intensity, low/high H&E color similarity). Across models and staining conditions, classification performance ranged from AUC 0.769-0.911 ($Δ$ = 0.142). Robustness ranged from 0.007-0.079 ($Δ$ = 0.072), and showed a weak inverse correlation with classification performance (Pearson r=-0.22, 95% CI [-0.34, -0.11]). Thus, we show that the proposed evaluation protocol enables robustness-informed CPath model selection and provides insight into performance shifts across H&E staining conditions, supporting the identification of operational ranges for reliable model deployment. Code is available at https://github.com/CTPLab/staining-robustness-evaluation .

Method: Formulates video-based seizure forecasting task using short pre-ictal video segments (3-10s) to predict seizures within 5s. Uses cross-species transfer learning framework leveraging large-scale rodent video data for auxiliary pretraining to capture seizure-related behavioral dynamics that generalize across species.

Result: Achieves over 70% prediction accuracy under strictly video-only setting and outperforms existing baselines.

Conclusion: Demonstrates potential of cross-species learning for building non-invasive, scalable early-warning systems for epilepsy using video data.

Abstract: Epileptic seizure forecasting is a clinically important yet challenging problem in epilepsy research. Existing approaches predominantly rely on neural signals such as electroencephalography (EEG), which require specialized equipment and limit long-term deployment in real-world settings. In contrast, video data provide a non-invasive and accessible alternative, yet existing video-based studies mainly focus on post-onset seizure detection, leaving seizure forecasting largely unexplored. In this work, we formulate a novel task of video-based epileptic seizure forecasting, where short pre-ictal video segments (3-10 seconds) are used to predict whether a seizure will occur within the subsequent 5 seconds. To address the scarcity of annotated human epilepsy videos, we propose a cross-species transfer learning framework that leverages large-scale rodent video data for auxiliary pretraining. This enables the model to capture seizure-related behavioral dynamics that generalize across species. Experimental results demonstrate that our approach achieves over 70% prediction accuracy under a strictly video-only setting and outperforms existing baselines. These findings highlight the potential of cross-species learning for building non-invasive, scalable early-warning systems for epilepsy.

Method: Proposes SG-NLF with: 1) hybrid representation based on spectral priors for smooth geometry reconstruction, 2) confidence-aware graph based on feature compatibility for global pose optimization, and 3) adversarial learning strategy for cross-frame consistency enhancement.

Result: SG-NLF improves reconstruction quality and pose accuracy by over 35.8% and 68.8% compared to previous state-of-the-art methods, especially effective in challenging low-frequency scenarios.

Conclusion: The framework provides a novel perspective for LiDAR view synthesis by successfully addressing pose-free reconstruction challenges through spectral-geometric integration and adversarial consistency learning.

Abstract: Neural Radiance Fields (NeRF) have shown remarkable success in image novel view synthesis (NVS), inspiring extensions to LiDAR NVS. However, most methods heavily rely on accurate camera poses for scene reconstruction. The sparsity and textureless nature of LiDAR data also present distinct challenges, leading to geometric holes and discontinuous surfaces. To address these issues, we propose SG-NLF, a pose-free LiDAR NeRF framework that integrates spectral information with geometric consistency. Specifically, we design a hybrid representation based on spectral priors to reconstruct smooth geometry. For pose optimization, we construct a confidence-aware graph based on feature compatibility to achieve global alignment. In addition, an adversarial learning strategy is introduced to enforce cross-frame consistency, thereby enhancing reconstruction quality. Comprehensive experiments demonstrate the effectiveness of our framework, especially in challenging low-frequency scenarios. Compared to previous state-of-the-art methods, SG-NLF improves reconstruction quality and pose accuracy by over 35.8% and 68.8%. Our work can provide a novel perspective for LiDAR view synthesis.

Method: Proposes VIRD with dual-axis transformation: 1) polar transformation on satellite view for horizontal correspondence, 2) context-enhanced positional attention on ground and polar-transformed satellite features to resolve vertical misalignment, 3) view-reconstruction loss to strengthen view invariance.

Result: Outperforms SOTA on KITTI and VIGOR datasets without orientation priors: reduces median position errors by 50.7% and 18.0%, and orientation errors by 76.5% and 46.8% respectively.

Conclusion: VIRD effectively bridges ground-satellite viewpoint gap through explicit dual-axis transformation and view-invariant learning, enabling accurate cross-view pose estimation for autonomous systems.

Abstract: Accurate global localization is crucial for autonomous driving and robotics, but GNSS-based approaches often degrade due to occlusion and multipath effects. As an emerging alternative, cross-view pose estimation predicts the 3-DoF camera pose corresponding to a ground-view image with respect to a geo-referenced satellite image. However, existing methods struggle to bridge the significant viewpoint gap between the ground and satellite views mainly due to limited spatial correspondences. We propose a novel cross-view pose estimation method that constructs view-invariant representations through dual-axis transformation (VIRD). VIRD first applies a polar transformation to the satellite view to establish horizontal correspondence, then uses context-enhanced positional attention on the ground and polar-transformed satellite features to resolve vertical misalignment, explicitly mitigating the viewpoint gap. A view-reconstruction loss is introduced to strengthen the view invariance further, encouraging the derived representations to reconstruct the original and cross-view images. Experiments on the KITTI and VIGOR datasets demonstrate that VIRD outperforms the state-of-the-art methods without orientation priors, reducing median position and orientation errors by 50.7% and 76.5% on KITTI, and 18.0% and 46.8% on VIGOR, respectively.

Method: Proposes IFDL-VLM pipeline that leverages location masks as explicit priors to guide vision-language models. These masks encode forgery concepts and help overcome VLM biases by providing additional context about manipulated regions, easing training optimization and enhancing interpretability.

Result: Achieves state-of-the-art performance on 9 popular benchmarks for both in-domain and cross-dataset generalization settings. Shows consistent improvements in detection, localization, and interpretability compared to previous methods.

Conclusion: The IFDL-VLM framework successfully addresses VLM biases in forgery detection by incorporating location masks as priors, demonstrating superior performance across multiple benchmarks and enhancing the interpretability of detection results.

Abstract: With the rapid rise of Artificial Intelligence Generated Content (AIGC), image manipulation has become increasingly accessible, posing significant challenges for image forgery detection and localization (IFDL). In this paper, we study how to fully leverage vision-language models (VLMs) to assist the IFDL task. In particular, we observe that priors from VLMs hardly benefit the detection and localization performance and even have negative effects due to their inherent biases toward semantic plausibility rather than authenticity. Additionally, the location masks explicitly encode the forgery concepts, which can serve as extra priors for VLMs to ease their training optimization, thus enhancing the interpretability of detection and localization results. Building on these findings, we propose a new IFDL pipeline named IFDL-VLM. To demonstrate the effectiveness of our method, we conduct experiments on 9 popular benchmarks and assess the model performance under both in-domain and cross-dataset generalization settings. The experimental results show that we consistently achieve new state-of-the-art performance in detection, localization, and interpretability.Code is available at: https://github.com/sha0fengGuo/IFDL-VLM.

Method: Three-part framework: 1) Created CameraMotionDataset with explicit camera control, 2) Diagnosed VideoLLM failures through probing experiments on vision encoders, 3) Proposed lightweight pipeline using 3D foundation models to extract geometric cues, predict motion primitives with temporal classifier, and inject via structured prompting.

Result: Substantial errors in camera motion recognition across off-the-shelf VideoLLMs; camera motion cues weakly represented in deeper ViT blocks; proposed injection method improves motion recognition and enables more camera-aware model responses.

Conclusion: Geometry-driven cue extraction and structured prompting are practical steps toward camera-aware VideoLLMs; the framework addresses a critical gap in multimodal understanding of fundamental geometric signals in video.

Abstract: Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\textbf{benchmarking}$, $\textbf{diagnosis}$, and $\textbf{injection}$. We curate $\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark–$\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at https://hf.co/datasets/fengyee/camera-motion-dataset-and-benchmark.

Method: SGMatch introduces a Semantic-Guided Local Cross-Attention module that integrates semantic features from vision foundation models into geometric descriptors while preserving local structural continuity. It also uses a regularization objective based on conditional flow matching that supervises a time-varying velocity field to encourage spatial smoothness of recovered correspondences.

Result: Experimental results on multiple benchmarks show SGMatch achieves competitive performance in near-isometric settings and consistent improvements under non-isometric deformations and topological noise.

Conclusion: SGMatch effectively addresses limitations of existing functional map pipelines by integrating semantic guidance and flow-based regularization, improving robustness to challenging non-isometric deformations and topological variations.

Abstract: Establishing accurate point-to-point correspondences between non-rigid 3D shapes remains a critical challenge, particularly under non-isometric deformations and topological noise. Existing functional map pipelines suffer from ambiguities that geometric descriptors alone cannot resolve, and spatial inconsistencies inherent in the projection of truncated spectral bases to dense pointwise correspondences. In this paper, we introduce SGMatch, a learning-based framework for semantic-guided non-rigid shape matching. Specifically, we design a Semantic-Guided Local Cross-Attention module that integrates semantic features from vision foundation models into geometric descriptors while preserving local structural continuity. Furthermore, we introduce a regularization objective based on conditional flow matching, which supervises a time-varying velocity field to encourage spatial smoothness of the recovered correspondences. Experimental results on multiple benchmarks demonstrate that SGMatch achieves competitive performance across near-isometric settings and consistent improvements under non-isometric deformations and topological noise.

Method: Proposes CleanSight, a training-free, plug-and-play defense that: (1) detects poisoned inputs based on relative visual-text attention ratio in cross-modal fusion layers, and (2) purifies inputs by selectively pruning suspicious high-attention visual tokens to neutralize backdoor activation.

Result: Extensive experiments show CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types, while preserving model utility on both clean and poisoned samples.

Conclusion: CleanSight provides an effective, efficient defense against backdoor attacks in LVLMs by leveraging the mechanistic understanding of attention stealing, offering a practical solution without retraining or performance degradation.

Abstract: Despite the strong multimodal performance, large vision-language models (LVLMs) are vulnerable during fine-tuning to backdoor attacks, where adversaries insert trigger-embedded samples into the training data to implant behaviors that can be maliciously activated at test time. Existing defenses typically rely on retraining backdoored parameters (e.g., adapters or LoRA modules) with clean data, which is computationally expensive and often degrades model performance. In this work, we provide a new mechanistic understanding of backdoor behaviors in LVLMs: the trigger does not influence prediction through low-level visual patterns, but through abnormal cross-modal attention redistribution, where trigger-bearing visual tokens steal attention away from the textual context - a phenomenon we term attention stealing. Motivated by this, we propose CleanSight, a training-free, plug-and-play defense that operates purely at test time. CleanSight (i) detects poisoned inputs based on the relative visual-text attention ratio in selected cross-modal fusion layers, and (ii) purifies the input by selectively pruning the suspicious high-attention visual tokens to neutralize the backdoor activation. Extensive experiments show that CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types, while preserving the model’s utility on both clean and poisoned samples.

Method: A training-free debiasing method requiring no annotated data that yields closed-form solutions in cross-modal space to jointly debias both visual and textual modalities across downstream tasks.

Result: Outperforms existing methods across diverse fairness metrics and datasets for both group and intersectional fairness in tasks like zero-shot image classification, text-to-image retrieval, and text-to-image generation while preserving task performance.

Conclusion: Proposes an effective debiasing approach for VLMs that achieves Pareto-optimal fairness with bounded utility losses without requiring training or annotated data.

Abstract: While Vision-Language Models (VLMs) have achieved remarkable performance across diverse downstream tasks, recent studies have shown that they can inherit social biases from the training data and further propagate them into downstream applications. To address this issue, various debiasing approaches have been proposed, yet most of them aim to improve fairness without having a theoretical guarantee that the utility of the model is preserved. In this paper, we introduce a debiasing method that yields a \textbf{closed-form} solution in the cross-modal space, achieving Pareto-optimal fairness with \textbf{bounded utility losses}. Our method is \textbf{training-free}, requires \textbf{no annotated data}, and can jointly debias both visual and textual modalities across downstream tasks. Extensive experiments show that our method outperforms existing methods in debiasing VLMs across diverse fairness metrics and datasets for both group and \textbf{intersectional} fairness in downstream tasks such as zero-shot image classification, text-to-image retrieval, and text-to-image generation while preserving task performance.

Method: Propose Mio (Multimodal Interactive Omni-Avatar) framework with five modules: Thinker (cognitive reasoning), Talker (speech generation), Face Animator, Body Animator, and Renderer. This unified architecture integrates reasoning with real-time multimodal embodiment.

Result: The framework achieves superior performance compared to state-of-the-art methods across all evaluated dimensions. A new benchmark is established for rigorous evaluation of interactive intelligence capabilities.

Conclusion: The contributions move digital humans beyond superficial imitation toward intelligent interaction, enabling personality-aligned expression, adaptive interaction, and self-evolution.

Abstract: We introduce Interactive Intelligence, a novel paradigm of digital human that is capable of personality-aligned expression, adaptive interaction, and self-evolution. To realize this, we present Mio (Multimodal Interactive Omni-Avatar), an end-to-end framework composed of five specialized modules: Thinker, Talker, Face Animator, Body Animator, and Renderer. This unified architecture integrates cognitive reasoning with real-time multimodal embodiment to enable fluid, consistent interaction. Furthermore, we establish a new benchmark to rigorously evaluate the capabilities of interactive intelligence. Extensive experiments demonstrate that our framework achieves superior performance compared to state-of-the-art methods across all evaluated dimensions. Together, these contributions move digital humans beyond superficial imitation toward intelligent interaction.

Method: Developed dots.mocr method that treats charts, diagrams, tables, and icons as first-class parsing targets, built comprehensive data engine from PDFs, webpages, and SVG assets, and trained a 3B-parameter model through staged pretraining and supervised fine-tuning.

Result: Ranks second only to Gemini 3 Pro on OCR Arena Elo leaderboard, surpasses existing open-source document parsing systems, achieves 83.9 on olmOCR Bench, and outperforms Gemini 3 Pro on image-to-SVG benchmarks for charts, UI layouts, scientific figures, and chemical diagrams.

Conclusion: MOCR paradigm enables joint parsing of text and graphics into unified representations, demonstrating scalable path for building large-scale image-to-code corpora for multimodal pretraining while preserving semantic relationships across document elements.

Abstract: We present Multimodal OCR (MOCR), a document parsing paradigm that jointly parses text and graphics into unified textual representations. Unlike conventional OCR systems that focus on text recognition and leave graphical regions as cropped pixels, our method, termed dots.mocr, treats visual elements such as charts, diagrams, tables, and icons as first-class parsing targets, enabling systems to parse documents while preserving semantic relationships across elements. It offers several advantages: (1) it reconstructs both text and graphics as structured outputs, enabling more faithful document reconstruction; (2) it supports end-to-end training over heterogeneous document elements, allowing models to exploit semantic relations between textual and visual components; and (3) it converts previously discarded graphics into reusable code-level supervision, unlocking multimodal supervision embedded in existing documents. To make this paradigm practical at scale, we build a comprehensive data engine from PDFs, rendered webpages, and native SVG assets, and train a compact 3B-parameter model through staged pretraining and supervised fine-tuning. We evaluate dots.mocr from two perspectives: document parsing and structured graphics parsing. On document parsing benchmarks, it ranks second only to Gemini 3 Pro on our OCR Arena Elo leaderboard, surpasses existing open-source document parsing systems, and sets a new state of the art of 83.9 on olmOCR Bench. On structured graphics parsing, dots.mocr achieves higher reconstruction quality than Gemini 3 Pro across image-to-SVG benchmarks, demonstrating strong performance on charts, UI layouts, scientific figures, and chemical diagrams. These results show a scalable path toward building large-scale image-to-code corpora for multimodal pretraining. Code and models are publicly available at https://github.com/rednote-hilab/dots.mocr.

Method: Proposes Visual-ERM, a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback by evaluating vision-to-code quality directly in the rendered visual space. Integrated into RL framework and benchmarked with VisualCritic-RewardBench.

Result: Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average). At 8B parameters, it outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models on VC-RewardBench.

Conclusion: Fine-grained visual reward supervision is both necessary and sufficient for vision-to-code reinforcement learning, regardless of task specificity. Visual-ERM provides effective multimodal feedback for structured visual data reconstruction tasks.

Abstract: Vision-to-code tasks require models to reconstruct structured visual inputs, such as charts, tables, and SVGs, into executable or structured representations with high visual fidelity. While recent Large Vision Language Models (LVLMs) achieve strong results via supervised fine-tuning, reinforcement learning remains challenging due to misaligned reward signals. Existing rewards either rely on textual rules or coarse visual embedding similarity, both of which fail to capture fine-grained visual discrepancies and are vulnerable to reward hacking. We propose Visual Equivalence Reward Model (Visual-ERM), a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback to evaluate vision-to-code quality directly in the rendered visual space. Integrated into RL, Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average), and further strengthens test-time scaling via reflection and revision. We also introduce VisualCritic-RewardBench (VC-RewardBench), a benchmark for judging fine-grained image-to-image discrepancies on structured visual data, where Visual-ERM at 8B decisively outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models. Our results suggest that fine-grained visual reward supervision is both necessary and sufficient for vision-to-code RL, regardless of task specificity.

Method: ESPIRE offers a simulated world that physically grounds VLMs and evaluates them on spatial-reasoning-centric robotic tasks. Tasks are decomposed into localization and execution components, both framed as generative problems rather than discriminative evaluations. The benchmark is systematically designed at both instruction and environment levels to ensure broad coverage of spatial reasoning scenarios.

Result: The paper uses ESPIRE to diagnose a range of frontier VLMs and provides in-depth analysis of their spatial reasoning behaviors, though specific results aren’t detailed in the abstract.

Conclusion: ESPIRE addresses limitations in current VLM evaluations by providing a comprehensive diagnostic benchmark for embodied spatial reasoning that enables fine-grained analysis beyond passive spatial reasoning toward reasoning to act.

Abstract: A recent trend in vision-language models (VLMs) has been to enhance their spatial cognition for embodied domains. Despite progress, existing evaluations have been limited both in paradigm and in coverage, hindering rapid, iterative model development. To address these limitations, we propose ESPIRE, a diagnostic benchmark for embodied spatial reasoning. ESPIRE offers a simulated world that physically grounds VLMs and evaluates them on spatial-reasoning-centric robotic tasks, thus narrowing the gap between evaluation and real-world deployment. To adapt VLMs to robotic tasks, we decompose each task into localization and execution, and frame both as generative problems, in stark contrast to predominant discriminative evaluations (e.g., via visual-question answering) that rely on distractors and discard execution. This decomposition further enables a fine-grained analysis beyond passive spatial reasoning toward reasoning to act. We systematically design ESPIRE both at the instruction level and at the environment level, ensuring broad coverage of spatial reasoning scenarios. We use ESPIRE to diagnose a range of frontier VLMs and provide in-depth analysis of their spatial reasoning behaviors.

Method: Developed automated data-curation pipeline to synthesize diverse topological anomalies with verifiable annotations. Created Topo-R1 framework with two-stage training: supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization (GRPO). Uses topology-aware composite reward with type-aware Hungarian matching, spatial localization scoring, and centerline Dice (clDice) reward.

Result: Topo-R1 establishes new paradigm for annotation-free topological quality assessment, consistently outperforming general-purpose VLMs and supervised baselines across all evaluation protocols.

Conclusion: The framework successfully endows VLMs with topology-aware perception for detecting topological anomalies without ground-truth supervision, addressing a critical gap in visual reasoning for medical imaging and other domains with tubular structures.

Abstract: Topological correctness is crucial for tubular structures such as blood vessels, nerve fibers, and road networks. Existing topology-preserving methods rely on domain-specific ground truth, which is costly and rarely transfers across domains. When deployed to a new domain without annotations, a key question arises: how can we detect topological anomalies without ground-truth supervision? We reframe this as topological anomaly detection, a structured visual reasoning task requiring a model to locate and classify topological errors in predicted segmentation masks. Vision-Language Models (VLMs) are natural candidates; however, we find that state-of-the-art VLMs perform nearly at random, lacking the fine-grained, topology-aware perception needed to identify sparse connectivity errors in dense structures. To bridge this gap, we develop an automated data-curation pipeline that synthesizes diverse topological anomalies with verifiable annotations across progressively difficult levels, thereby constructing the first large-scale, multi-domain benchmark for this task. We then introduce Topo-R1, a framework that endows VLMs with topology-aware perception via two-stage training: supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization (GRPO). Central to our approach is a topology-aware composite reward that integrates type-aware Hungarian matching for structured error classification, spatial localization scoring, and a centerline Dice (clDice) reward that directly penalizes connectivity disruptions, thereby jointly incentivizing semantic precision and structural fidelity. Extensive experiments demonstrate that Topo-R1 establishes a new paradigm for annotation-free topological quality assessment, consistently outperforming general-purpose VLMs and supervised baselines across all evaluation protocols.

Method: Proposes VTON-IQA framework with: 1) VTON-QBench dataset (62,688 try-on images from 14 models, 431,800 human annotations), 2) Interleaved Cross-Attention module that inserts cross-attention between self-attention and MLP in transformer blocks to model garment-person interactions, 3) Reference-free quality prediction without ground-truth images.

Result: VTON-IQA achieves reliable human-aligned image-level quality prediction. The framework enables comprehensive benchmark evaluation of 14 representative VTON models, providing a robust assessment tool for virtual try-on systems.

Conclusion: VTON-IQA addresses critical evaluation challenges in virtual try-on systems by providing a reference-free, human-aligned quality assessment framework that explicitly models garment-person interactions, enabling reliable evaluation of individual generated images without ground-truth references.

Abstract: Given a person image and a garment image, image-based Virtual Try-ON (VTON) synthesizes a try-on image of the person wearing the target garment. As VTON systems become increasingly important in practical applications such as fashion e-commerce, reliable evaluation of their outputs has emerged as a critical challenge. In real-world scenarios, ground-truth images of the same person wearing the target garment are typically unavailable, making reference-based evaluation impractical. Moreover, widely used distribution-level metrics such as Fréchet Inception Distance and Kernel Inception Distance measure dataset-level similarity and fail to reflect the perceptual quality of individual generated images. To address these limitations, we propose Image Quality Assessment for Virtual Try-On (VTON-IQA), a reference-free framework for human-aligned, image-level quality assessment without requiring ground-truth images. To model human perceptual judgments, we construct VTON-QBench, a large-scale human-annotated benchmark comprising 62,688 try-on images generated by 14 representative VTON models and 431,800 quality annotations collected from 13,838 qualified annotators. To the best of our knowledge, this is the largest dataset to date for human subjective evaluation in virtual try-on. Evaluating virtual try-on quality requires verifying both garment fidelity and the preservation of person-specific details. To explicitly model such interactions, we introduce an Interleaved Cross-Attention module that extends standard transformer blocks by inserting a cross-attention layer between self-attention and MLP in the latter blocks. Extensive experiments show that VTON-IQA achieves reliable human-aligned image-level quality prediction. Moreover, we conduct a comprehensive benchmark evaluation of 14 representative VTON models using VTON-IQA.

Method: 1. Region-Aware Prompt Augmentation (RAPTA): Uses object detector to find salient regions and create semantically grounded prompt variants, randomly sampled during training to increase diversity while maintaining semantic alignment. 2. Attention-Driven Multimodal Copy Detection (ADMCD): Aggregates local patch, global semantic, and texture cues with lightweight transformer to produce fused representation, applying simple thresholded decision rules to detect copying without large annotated datasets.

Result: RAPTA reduces overfitting while maintaining high synthesis quality. ADMCD reliably detects copying and outperforms single-modal metrics.

Conclusion: The proposed complementary methods effectively address memorization issues in text-to-image diffusion models - RAPTA prevents overfitting during training while ADMCD detects copying at inference time, together providing a comprehensive solution to copyright and privacy concerns.

Abstract: State-of-the-art text-to-image diffusion models can produce impressive visuals but may memorize and reproduce training images, creating copyright and privacy risks. Existing prompt perturbations applied at inference time, such as random token insertion or embedding noise, may lower copying but often harm image-prompt alignment and overall fidelity. To address this, we introduce two complementary methods. First, Region-Aware Prompt Augmentation (RAPTA) uses an object detector to find salient regions and turn them into semantically grounded prompt variants, which are randomly sampled during training to increase diversity, while maintaining semantic alignment. Second, Attention-Driven Multimodal Copy Detection (ADMCD) aggregates local patch, global semantic, and texture cues with a lightweight transformer to produce a fused representation, and applies simple thresholded decision rules to detect copying without training with large annotated datasets. Experiments show that RAPTA reduces overfitting while maintaining high synthesis quality, and that ADMCD reliably detects copying, outperforming single-modal metrics.

Method: Developed a learning-from-observation framework using wind-tunnel PIV data, comparing Kriging interpolation with three deep learning models (UNet, Vision Transformer Autoencoder, Conditional Wasserstein GAN). Evaluated single vs. mixed wind-direction training strategies, sensor densities (5-30), robustness to sensor perturbations (±1 grid), and optimized sensor placement using Proper Orthogonal Decomposition with QR decomposition.

Result: Deep learning methods significantly outperformed Kriging interpolation (up to 32.7% SSIM improvement, 24.2% FAC2, 27.8% NMSE). Mixed wind-direction training provided substantial gains (up to 173.7% SSIM, 16.7% FAC2, 98.3% MG). QR-based optimization improved robustness by up to 27.8% under sensor perturbations, though with metric-dependent trade-offs.

Conclusion: Deep learning effectively reconstructs rooftop wind fields from sparse sensors, with performance enhanced by mixed wind-direction training and optimized sensor placement. Joint consideration of sensor configuration, optimization, and training strategy is essential for reliable deployment, and experimental data provides practical guidance for real-world applications.

Abstract: Real-time rooftop wind-speed distribution is important for the safe operation of drones and urban air mobility systems, wind control systems, and rooftop utilization. However, rooftop flows show strong nonlinearity, separation, and cross-direction variability, which make flow field reconstruction from sparse sensors difficult. This study develops a learning-from-observation framework using wind-tunnel experimental data obtained by Particle Image Velocimetry (PIV) and compares Kriging interpolation with three deep learning models: UNet, Vision Transformer Autoencoder (ViTAE), and Conditional Wasserstein GAN (CWGAN). We evaluate two training strategies, single wind-direction training (SDT) and mixed wind-direction training (MDT), across sensor densities from 5 to 30, test robustness under sensor position perturbations of plus or minus 1 grid, and optimize sensor placement via Proper Orthogonal Decomposition with QR decomposition. Results show that deep learning methods can reconstruct rooftop wind fields from sparse sensor data effectively. Compared with Kriging interpolation, the deep learning models improved SSIM by up to 32.7%, FAC2 by 24.2%, and NMSE by 27.8%. Mixed wind-direction training further improved performance, with gains of up to 173.7% in SSIM, 16.7% in FAC2, and 98.3% in MG compared with single-direction training. The results also show that sensor configuration, optimization, and training strategy should be considered jointly for reliable deployment. QR-based optimization improved robustness by up to 27.8% under sensor perturbations, although with metric-dependent trade-offs. Training on experimental rather than simulated data also provides practical guidance for method selection and sensor placement in different scenarios.

Method: Reinterprets image restoration as a progressive generative process rather than static regression. Leverages pretrained video models to simulate gradual refinement from degraded inputs to high-fidelity outputs. Uses only 1,000 multi-task training samples (less than 2% of existing methods) to adapt video models for restoration tasks.

Result: Pretrained video models can perform competitive image restoration with minimal adaptation, achieving multiple tasks with a single model, rivaling specialized architectures designed explicitly for restoration purposes.

Conclusion: Video generative models implicitly learn powerful and transferable restoration priors that can be activated with minimal data, challenging traditional boundaries between generative modeling and low-level vision, and opening new design paradigms for foundation models in visual tasks.

Abstract: Large-scale video generative models are trained on vast and diverse visual data, enabling them to internalize rich structural, semantic, and dynamic priors of the visual world. While these models have demonstrated impressive generative capability, their potential as general-purpose visual learners remains largely untapped. In this work, we introduce V-Bridge, a framework that bridges this latent capacity to versatile few-shot image restoration tasks. We reinterpret image restoration not as a static regression problem, but as a progressive generative process, and leverage video models to simulate the gradual refinement from degraded inputs to high-fidelity outputs. Surprisingly, with only 1,000 multi-task training samples (less than 2% of existing restoration methods), pretrained video models can be induced to perform competitive image restoration, achieving multiple tasks with a single model, rivaling specialized architectures designed explicitly for this purpose. Our findings reveal that video generative models implicitly learn powerful and transferable restoration priors that can be activated with only extremely limited data, challenging the traditional boundary between generative modeling and low-level vision, and opening a new design paradigm for foundation models in visual tasks.

Method: Introduces a structured evaluation taxonomy for abstractive spatiotemporal reasoning and constructs a controllable, scenario-driven synthetic egocentric video dataset. Creates VAEX-BENCH with five abstractive reasoning tasks and their extractive counterparts across object-, room-, and floor-plan-level scenarios.

Result: Extensive experiments compare state-of-the-art MLLMs under extractive and abstractive settings, exposing their limitations on abstractive tasks and providing fine-grained analysis of underlying bottlenecks.

Conclusion: The benchmark reveals significant gaps in MLLMs’ abstractive spatiotemporal reasoning capabilities and provides a structured framework for future research in this area.

Abstract: The growing interest in embodied agents increases the demand for spatiotemporal video understanding, yet existing benchmarks largely emphasize extractive reasoning, where answers can be explicitly presented within spatiotemporal events. It remains unclear whether multimodal large language models can instead perform abstractive spatiotemporal reasoning, which requires integrating observations over time, combining dispersed cues, and inferring implicit spatial and contextual structure. To address this gap, we formalize abstractive spatiotemporal reasoning from videos by introducing a structured evaluation taxonomy that systematically targets its core dimensions and construct a controllable, scenario-driven synthetic egocentric video dataset tailored to evaluate abstractive spatiotemporal reasoning capabilities, spanning object-, room-, and floor-plan-level scenarios. Based on this framework, we present VAEX-BENCH, a benchmark comprising five abstractive reasoning tasks together with their extractive counterparts. Our extensive experiments compare the performance of state-of-the-art MLLMs under extractive and abstractive settings, exposing their limitations on abstractive tasks and providing a fine-grained analysis of the underlying bottlenecks. The dataset will be released soon.

Method: Proposes a taxonomy of manufacturability metrics along configuration dependence and measurement type axes, and introduces BenDFM - a synthetic dataset of 20,000 sheet metal bending parts (both manufacturable/unmanufacturable) generated with process-aware simulations, with multiple manufacturability labels. Benchmarks two state-of-the-art 3D learning architectures on this dataset.

Result: Graph-based representations capturing relationships between part surfaces achieve better accuracy, while predicting metrics dependent on specific manufacturing setups remains challenging. The BenDFM dataset enables systematic study of learning-based DFM challenges.

Conclusion: The proposed taxonomy and BenDFM dataset provide a foundation for advancing learning-based manufacturability prediction, particularly for sheet metal bending, with graph-based architectures showing promise for capturing geometric relationships.

Abstract: Predicting the manufacturability of CAD designs early, in terms of both feasibility and required effort, is a key goal of Design for Manufacturing (DFM). Despite advances in deep learning for CAD and its widespread use in manufacturing process selection, learning-based approaches for predicting manufacturability within a specific process remain limited. Two key challenges limit progress: inconsistency across prior work in how manufacturability is defined and consequently in the associated learning targets, and a scarcity of suitable datasets. Existing labels vary significantly: they may reflect intrinsic design constraints or depend on specific manufacturing capabilities (such as available tools), and they range from discrete feasibility checks to continuous complexity measures. Furthermore, industrial datasets typically contain only manufacturable parts, offering little signal for infeasible cases, while existing synthetic datasets focus on simple geometries and subtractive processes. To address these gaps, we propose a taxonomy of manufacturability metrics along the axes of configuration dependence and measurement type, allowing clearer scoping of generalizability and learning objectives. Next, we introduce BenDFM, the first synthetic dataset for manufacturability assessment in sheet metal bending. BenDFM contains 20,000 parts, both manufacturable and unmanufacturable, generated with process-aware bending simulations, providing both folded and unfolded geometries and multiple manufacturability labels across the taxonomy, enabling systematic study of previously unexplored learning-based DFM challenges. We benchmark two state-of-the-art 3D learning architectures on BenDFM, showing that graph-based representations that capture relationships between part surfaces achieve better accuracy, and that predicting metrics that depend on specific manufacturing setups remains more challenging.

Method: NOIR embeds discrete medical signals into shared Implicit Neural Representations and learns a Neural Operator that maps between their latent modulations, enabling resolution-independent function-to-function transformations.

Result: Achieves competitive performance at native resolution across multiple 2D/3D tasks (segmentation, shape completion, image-to-image translation, synthesis) on public and clinical datasets, while demonstrating robustness to unseen discretizations and satisfying theoretical properties of neural operators.

Conclusion: NOIR provides a novel continuous function space approach to medical imaging that outperforms discrete grid-based methods in resolution independence and robustness, opening new directions for medical image analysis.

Abstract: This paper presents NOIR, a framework that reframes core medical imaging tasks as operator learning between continuous function spaces, challenging the prevailing paradigm of discrete grid-based deep learning. Instead of operating on fixed pixel or voxel grids, NOIR embeds discrete medical signals into shared Implicit Neural Representations and learns a Neural Operator that maps between their latent modulations, enabling resolution-independent function-to-function transformations. We evaluate NOIR across multiple 2D and 3D downstream tasks, including segmentation, shape completion, image-to-image translation, and image synthesis, on several public datasets such as Shenzhen, OASIS-4, SkullBreak, fastMRI, as well as an in-house clinical dataset. It achieves competitive performance at native resolution while demonstrating strong robustness to unseen discretizations, and empirically satisfies key theoretical properties of neural operators. The project page is available here: https://github.com/Sidaty1/NOIR-io.

Method: Developed FDeID-Toolbox with modular architecture: (1) standardized data loaders for benchmark datasets, (2) unified implementations of methods from classical approaches to state-of-the-art generative models, (3) flexible inference pipelines, and (4) systematic evaluation protocols covering privacy, utility, and quality metrics.

Result: The toolbox enables fair and reproducible comparison of diverse FDeID methods under consistent conditions, addressing the field’s fragmentation and evaluation inconsistencies.

Conclusion: FDeID-Toolbox provides a comprehensive solution for reproducible face de-identification research, facilitating standardized evaluation and comparison across methods while addressing the field’s fragmentation challenges.

Abstract: Face de-identification (FDeID) aims to remove personally identifiable information from facial images while preserving task-relevant utility attributes such as age, gender, and expression. It is critical for privacy-preserving computer vision, yet the field suffers from fragmented implementations, inconsistent evaluation protocols, and incomparable results across studies. These challenges stem from the inherent complexity of the task: FDeID spans multiple downstream applications (e.g., age estimation, gender recognition, expression analysis) and requires evaluation across three dimensions (e.g., privacy protection, utility preservation, and visual quality), making existing codebases difficult to use and extend. To address these issues, we present FDeID-Toolbox, a comprehensive toolbox designed for reproducible FDeID research. Our toolbox features a modular architecture comprising four core components: (1) standardized data loaders for mainstream benchmark datasets, (2) unified method implementations spanning classical approaches to SOTA generative models, (3) flexible inference pipelines, and (4) systematic evaluation protocols covering privacy, utility, and quality metrics. Through experiments, we demonstrate that FDeID-Toolbox enables fair and reproducible comparison of diverse FDeID methods under consistent conditions.

Method: f-CBM uses two complementary strategies: 1) a differentiable leakage loss to mitigate leakage (where concept representations encode extraneous task-relevant information), and 2) a Kolmogorov-Arnold Network prediction head that provides sufficient expressiveness to improve concept detection. The framework is built on a vision-language backbone and applies to both image-text and text-only datasets.

Result: Experiments demonstrate that f-CBM achieves the best trade-off between task accuracy, concept detection, and leakage reduction. The framework applies seamlessly to both image and text or text-only datasets, making it versatile across modalities.

Conclusion: f-CBM provides a faithful multimodal CBM framework that jointly addresses concept detection and leakage mitigation, offering improved interpretability while maintaining predictive performance across different modalities.

Abstract: Concept Bottleneck Models (CBMs) are interpretable models that route predictions through a layer of human-interpretable concepts. While widely studied in vision and, more recently, in NLP, CBMs remain largely unexplored in multimodal settings. For their explanations to be faithful, CBMs must satisfy two conditions: concepts must be properly detected, and concept representations must encode only their intended semantics, without smuggling extraneous task-relevant or inter-concept information into final predictions, a phenomenon known as leakage. Existing approaches treat concept detection and leakage mitigation as separate problems, and typically improve one at the expense of predictive accuracy. In this work, we introduce f-CBM, a faithful multimodal CBM framework built on a vision-language backbone that jointly targets both aspects through two complementary strategies: a differentiable leakage loss to mitigate leakage, and a Kolmogorov-Arnold Network prediction head that provides sufficient expressiveness to improve concept detection. Experiments demonstrate that f-CBM achieves the best trade-off between task accuracy, concept detection, and leakage reduction, while applying seamlessly to both image and text or text-only datasets, making it versatile across modalities.

Method: Proposes a novel lightweight perception scheduling framework that leverages outputs from previous frames to estimate and schedule necessary perception modules in real-time based on scene context, inspired by the Relevance concept and information sparsity in HRC events.

Result: Reduces computational latency by up to 27.52% compared to conventional parallel perception pipelines, achieves 72.73% improvement in MMPose activation recall, and maintains high keyframe accuracy up to 98%.

Conclusion: The framework effectively enhances real-time perception efficiency without significantly compromising accuracy, showing potential as a scalable and systematic solution for multimodal streaming perception systems in HRC.

Abstract: In modern human-robot collaboration (HRC) applications, multiple perception modules jointly extract visual, auditory, and contextual cues to achieve comprehensive scene understanding, enabling the robot to provide appropriate assistance to human agents intelligently. While executing multiple perception modules on a frame-by-frame basis enhances perception quality in offline settings, it inevitably accumulates latency, leading to a substantial decline in system performance in streaming perception scenarios. Recent work in scene understanding, termed Relevance, has established a solid foundation for developing efficient methodologies in HRC. However, modern perception pipelines still face challenges related to information redundancy and suboptimal allocation of computational resources. Drawing inspiration from the Relevance concept and the information sparsity in HRC events, we propose a novel lightweight perception scheduling framework that efficiently leverages output from previous frames to estimate and schedule necessary perception modules in real-time based on scene context. The experimental results demonstrate that the proposed perception scheduling framework effectively reduces computational latency by up to 27.52% compared to conventional parallel perception pipelines, while also achieving a 72.73% improvement in MMPose activation recall. Additionally, the framework demonstrates high keyframe accuracy, achieving rates of up to 98%. The results validate the framework’s capability to enhance real-time perception efficiency without significantly compromising accuracy. The framework shows potential as a scalable and systematic solution for multimodal streaming perception systems in HRC.

Method: Three-stage approach: 1) Preprocess MRI images and extract compact NNMF feature representations, selecting discriminative components using statistical metrics; 2) Train lightweight CNN classifier on selected features; 3) Add diffusion-based feature-space purification module with forward noise and learned denoiser network before classification.

Result: The framework achieves competitive classification performance while significantly enhancing robustness against adversarial attacks (evaluated using AutoAttack), maintaining both clean accuracy and robust accuracy.

Conclusion: Combining interpretable NNMF-based representations with lightweight deep learning and diffusion-based defense provides an effective, reliable solution for medical image classification under adversarial conditions.

Abstract: Brain tumor classification from magnetic resonance imaging, which is also known as MRI, plays a sensitive role in computer-assisted diagnosis systems. In recent years, deep learning models have achieved high classification accuracy. However, their sensitivity to adversarial perturbations has become an important reliability concern in medical applications. This study suggests a robust brain tumor classification framework that combines Non-Negative Matrix Factorization (NNMF or NMF), lightweight convolutional neural networks (CNNs), and diffusion-based feature purification. Initially, MRI images are preprocessed and converted into a non-negative data matrix, from which compact and interpretable NNMF feature representations are extracted. Statistical metrics, including AUC, Cohen’s d, and p-values, are used to rank and choose the most discriminative components. Then, a lightweight CNN classifier is trained directly on the selected feature groups. To improve adversarial robustness, a diffusion-based feature-space purification module is introduced. A forward noise method followed by a learned denoiser network is used before classification. System performance is estimated using both clean accuracy and robust accuracy under powerful adversarial attacks created by AutoAttack. The experimental results show that the proposed framework achieves competitive classification performance while significantly enhancing robustness against adversarial perturbations.The findings presuppose that combining interpretable NNMF-based representations with a lightweight deep approach and diffusion-based defense technique supplies an effective and reliable solution for medical image classification under adversarial conditions.

Method: 1) Introduces ActionGenome4D dataset with 4D scenes via 3D reconstruction, world-frame bounding boxes, and dense relationship annotations. 2) Formalizes World Scene Graph Generation (WSGG) task. 3) Proposes three methods: PWG (zero-order feature buffer), MWAE (masked completion with cross-view retrieval), and 4DST (temporal attention with 3D motion features). 4) Evaluates VLMs using Graph RAG approaches.

Result: The paper establishes baselines for unlocalized relationship prediction and advances video scene understanding toward world-centric, temporally persistent reasoning that handles both observed and unobserved objects.

Conclusion: WSGG advances video scene understanding toward world-centric, temporally persistent, and interpretable scene reasoning by addressing limitations of frame-centric approaches and enabling reasoning about occluded objects.

Abstract: Spatio-temporal scene graphs provide a principled representation for modeling evolving object interactions, yet existing methods remain fundamentally frame-centric: they reason only about currently visible objects, discard entities upon occlusion, and operate in 2D. To address this, we first introduce ActionGenome4D, a dataset that upgrades Action Genome videos into 4D scenes via feed-forward 3D reconstruction, world-frame oriented bounding boxes for every object involved in actions, and dense relationship annotations including for objects that are temporarily unobserved due to occlusion or camera motion. Building on this data, we formalize World Scene Graph Generation (WSGG), the task of constructing a world scene graph at each timestamp that encompasses all interacting objects in the scene, both observed and unobserved. We then propose three complementary methods, each exploring a different inductive bias for reasoning about unobserved objects: PWG (Persistent World Graph), which implements object permanence via a zero-order feature buffer; MWAE (Masked World Auto-Encoder), which reframes unobserved-object reasoning as masked completion with cross-view associative retrieval; and 4DST (4D Scene Transformer), which replaces the static buffer with differentiable per-object temporal attention enriched by 3D motion and camera-pose features. We further design and evaluate the performance of strong open-source Vision-Language Models on the WSGG task via a suite of Graph RAG-based approaches, establishing baselines for unlocalized relationship prediction. WSGG thus advances video scene understanding toward world-centric, temporally persistent, and interpretable scene reasoning.

Method: STEVO-Bench applies observation control to evolving processes via three types of instructions: occluder insertion, turning off lights, and specifying camera “lookaway” trajectories. It evaluates video models with and without camera control across diverse natural evolutions.

Result: The benchmark exposes limitations in video world models’ ability to decouple state evolution from observation. Analysis reveals data and architecture biases in present-day models, showing they struggle with generating plausible evolutions when observation is controlled.

Conclusion: STEVO-Bench provides a systematic evaluation protocol to detect and disentangle failure modes in video world models, offering new insights into their biases and limitations regarding state evolution independent of observation.

Abstract: Evolutions in the world, such as water pouring or ice melting, happen regardless of being observed. Video world models generate “worlds” via 2D frame observations. Can these generated “worlds” evolve regardless of observation? To probe this question, we design a benchmark to evaluate whether video world models can decouple state evolution from observation. Our benchmark, STEVO-Bench, applies observation control to evolving processes via instructions of occluder insertion, turning off the light, or specifying camera “lookaway” trajectories. By evaluating video models with and without camera control for a diverse set of naturally-occurring evolutions, we expose their limitations in decoupling state evolution from observation. STEVO-Bench proposes an evaluation protocol to automatically detect and disentangle failure modes of video world models across key aspects of natural state evolution. Analysis of STEVO-Bench results provide new insight into potential data and architecture bias of present-day video world models. Project website: https://glab-caltech.github.io/STEVOBench/. Blog: https://ziqi-ma.github.io/blog/2026/outofsight/

Method: Latent diffusion model combining variational autoencoder for dimension reduction and U-net for denoising. Applied to conditional 2D three-facies (channel-levee-mud) systems. Includes stability tests and ensemble-based data assimilation.

Result: Generated realizations visually consistent with geomodeling software. Quantitative metrics show agreement in spatial and flow-response statistics. Successful uncertainty reduction and consistent posterior geomodels in data assimilation tests.

Conclusion: Latent diffusion models provide effective geological parameterization for data assimilation, maintaining geological realism while enabling uncertainty quantification.

Abstract: Geological parameterization entails the representation of a geomodel using a small set of latent variables and a mapping from these variables to grid-block properties such as porosity and permeability. Parameterization is useful for data assimilation (history matching), as it maintains geological realism while reducing the number of variables to be determined. Diffusion models are a new class of generative deep-learning procedures that have been shown to outperform previous methods, such as generative adversarial networks, for image generation tasks. Diffusion models are trained to “denoise”, which enables them to generate new geological realizations from input fields characterized by random noise. Latent diffusion models, which are the specific variant considered in this study, provide dimension reduction through use of a low-dimensional latent variable. The model developed in this work includes a variational autoencoder for dimension reduction and a U-net for the denoising process. Our application involves conditional 2D three-facies (channel-levee-mud) systems. The latent diffusion model is shown to provide realizations that are visually consistent with samples from geomodeling software. Quantitative metrics involving spatial and flow-response statistics are evaluated, and general agreement between the diffusion-generated models and reference realizations is observed. Stability tests are performed to assess the smoothness of the parameterization method. The latent diffusion model is then used for ensemble-based data assimilation. Two synthetic “true” models are considered. Significant uncertainty reduction, posterior P${10}$-P${90}$ forecasts that generally bracket observed data, and consistent posterior geomodels, are achieved in both cases. PLEASE CITE AS: 10.1016/j.cageo.2024.105755 https://www.sciencedirect.com/science/article/pii/S0098300424002383 NOT WITH THE ARXIV VERSION

Method: Two-stage framework: 1) Motion reasoning stage using instance flow (sparse-to-dense motion representation) to integrate partial user-defined motions and motion inpainting to reason about motions of other objects, 2) Visual synthesis stage to generate final videos.

Result: Motion Dreamer significantly outperforms existing methods, achieving superior motion plausibility and visual realism in boundary conditional motion reasoning tasks.

Conclusion: The framework bridges the gap towards practical boundary conditional motion reasoning by effectively handling partial motion constraints while maintaining physical consistency and visual quality.

Abstract: Recent advances in video generation have shown promise for generating future scenarios, critical for planning and control in autonomous driving and embodied intelligence. However, real-world applications demand more than visually plausible predictions; they require reasoning about object motions based on explicitly defined boundary conditions, such as initial scene image and partial object motion. We term this capability Boundary Conditional Motion Reasoning. Current approaches either neglect explicit user-defined motion constraints, producing physically inconsistent motions, or conversely demand complete motion inputs, which are rarely available in practice. Here we introduce Motion Dreamer, a two-stage framework that explicitly separates motion reasoning from visual synthesis, addressing these limitations. Our approach introduces instance flow, a sparse-to-dense motion representation enabling effective integration of partial user-defined motions, and the motion inpainting strategy to robustly enable reasoning motions of other objects. Extensive experiments demonstrate that Motion Dreamer significantly outperforms existing methods, achieving superior motion plausibility and visual realism, thus bridging the gap towards practical boundary conditional motion reasoning. Our webpage is available: https://envision-research.github.io/MotionDreamer/.

Method: Develops a theoretical framework analyzing positional encoding in terms of shifted basis functions, establishes conditions for effective embeddings based on stable rank and distance preservation trade-offs, and empirically verifies that complex positional encodings can replace network depth.

Result: Shows that alternative non-Fourier embedding functions can work effectively, demonstrates that Fourier features are a special case of their general theory, and proves that trading positional embedding complexity for network depth is orders of magnitude faster than current approaches.

Conclusion: Provides a more general theory for positional encodings beyond Fourier analysis, showing that complex positional embeddings can reduce the need for deep networks while maintaining performance, offering significant speed improvements.

Abstract: It is well noted that coordinate-based MLPs benefit – in terms of preserving high-frequency information – through the encoding of coordinate positions as an array of Fourier features. Hitherto, the rationale for the effectiveness of these positional encodings has been mainly studied through a Fourier lens. In this paper, we strive to broaden this understanding by showing that alternative non-Fourier embedding functions can indeed be used for positional encoding. Moreover, we show that their performance is entirely determined by a trade-off between the stable rank of the embedded matrix and the distance preservation between embedded coordinates. We further establish that the now ubiquitous Fourier feature mapping of position is a special case that fulfills these conditions. Consequently, we present a more general theory to analyze positional encoding in terms of shifted basis functions. In addition, we argue that employing a more complex positional encoding – that scales exponentially with the number of modes – requires only a linear (rather than deep) coordinate function to achieve comparable performance. Counter-intuitively, we demonstrate that trading positional embedding complexity for network deepness is orders of magnitude faster than current state-of-the-art; despite the additional embedding complexity. To this end, we develop the necessary theoretical formulae and empirically verify that our theoretical claims hold in practice.

Method: Theoretical analysis delving into the intricate relationship between network initialization and activation functions, examining how these factors influence optimization dynamics in Neural Fields.

Result: Reveals deep-seated connections among network initialization, architectural choices, and optimization processes, emphasizing the need for holistic design approaches.

Conclusion: Provides foundational theoretical insights for robust optimization of Neural Fields, addressing a critical gap in the field’s theoretical understanding.

Abstract: In the realm of computer vision, Neural Fields have gained prominence as a contemporary tool harnessing neural networks for signal representation. Despite the remarkable progress in adapting these networks to solve a variety of problems, the field still lacks a comprehensive theoretical framework. This article aims to address this gap by delving into the intricate interplay between initialization and activation, providing a foundational basis for the robust optimization of Neural Fields. Our theoretical insights reveal a deep-seated connection among network initialization, architectural choices, and the optimization process, emphasizing the need for a holistic approach when designing cutting-edge Neural Fields.

Method: The authors introduce weight conditioning, a novel normalization technique that narrows the gap between the smallest and largest singular values of weight matrices. This approach results in better-conditioned matrices. The method is theoretically shown to smooth the loss landscape, thereby improving convergence of stochastic gradient descent algorithms.

Result: Empirical validation across various neural network architectures including CNNs, Vision Transformers (ViT), Neural Radiance Fields (NeRF), and 3D shape modeling shows that the proposed normalization method is competitive with and often outperforms existing weight normalization techniques from the literature.

Conclusion: Weight conditioning is an effective normalization technique that improves matrix conditioning, smooths loss landscapes, and enhances convergence across diverse neural network architectures, offering advantages over existing normalization methods.

Abstract: In this article, we introduce a novel normalization technique for neural network weight matrices, which we term weight conditioning. This approach aims to narrow the gap between the smallest and largest singular values of the weight matrices, resulting in better-conditioned matrices. The inspiration for this technique partially derives from numerical linear algebra, where well-conditioned matrices are known to facilitate stronger convergence results for iterative solvers. We provide a theoretical foundation demonstrating that our normalization technique smoothens the loss landscape, thereby enhancing convergence of stochastic gradient descent algorithms. Empirically, we validate our normalization across various neural network architectures, including Convolutional Neural Networks (CNNs), Vision Transformers (ViT), Neural Radiance Fields (NeRF), and 3D shape modeling. Our findings indicate that our normalization method is not only competitive but also outperforms existing weight normalization techniques from the literature.

Method: Proposes EEG-VJEPA, adapting Video Joint Embedding Predictive Architecture (V-JEPA) for EEG classification. Treats EEG as video-like sequences, learns spatiotemporal representations using joint embeddings and adaptive masking. First work to apply V-JEPA to EEG classification.

Result: Outperforms existing state-of-the-art models on Temple University Hospital Abnormal EEG dataset. Captures physiologically relevant spatial and temporal patterns, offers interpretable embeddings for human-AI collaboration in diagnostics.

Conclusion: EEG-VJEPA is a promising framework for scalable, trustworthy EEG analysis in clinical settings, with potential for human-AI collaboration in diagnostic workflows through interpretable embeddings.

Abstract: EEG signals capture brain activity with high temporal and low spatial resolution, supporting applications such as neurological diagnosis, cognitive monitoring, and brain-computer interfaces. However, effective analysis is hindered by limited labeled data, high dimensionality, and the absence of scalable models that fully capture spatiotemporal dependencies. Existing self-supervised learning (SSL) methods often focus on either spatial or temporal features, leading to suboptimal representations. To this end, we propose EEG-VJEPA, a novel adaptation of the Video Joint Embedding Predictive Architecture (V-JEPA) for EEG classification. By treating EEG as video-like sequences, EEG-VJEPA learns semantically meaningful spatiotemporal representations using joint embeddings and adaptive masking. To our knowledge, this is the first work that exploits V-JEPA for EEG classification and explores the visual concepts learned by the model. Evaluations on the publicly available Temple University Hospital (TUH) Abnormal EEG dataset show that EEG-VJEPA outperforms existing state-of-the-art models in classification accuracy. Beyond classification accuracy, EEG-VJEPA captures physiologically relevant spatial and temporal signal patterns, offering interpretable embeddings that may support human-AI collaboration in diagnostic workflows. These findings position EEG-VJEPA as a promising framework for scalable, trustworthy EEG analysis in real-world clinical settings.

Method: Introduces 3DGS-DET with two key innovations: (1) 2D Boundary Guidance to enhance spatial distribution clarity by leveraging 2D image information, and (2) Box-Focused Sampling using 2D boxes to generate object probability distributions in 3D space for effective probabilistic sampling that retains object blobs while reducing background noise.

Result: Significantly outperforms state-of-the-art NeRF-based method NeRF-Det++, achieving +6.0 mAP@0.25 and +7.8 mAP@0.5 improvements on ScanNet, and +14.9 mAP@0.25 improvement on ARKITScenes.

Conclusion: 3D Gaussian Splatting can be effectively adapted for indoor 3D object detection by addressing representation challenges through 2D boundary guidance and box-focused sampling, demonstrating superior performance over NeRF-based approaches.

Abstract: Neural Radiance Fields (NeRF) have been adapted for indoor 3D Object Detection (3DOD), offering a promising approach to indoor 3DOD via view-synthesis representation. But its implicit nature limits representational capacity. Recently, 3D Gaussian Splatting (3DGS) has emerged as an explicit 3D representation that addresses the limitation. This work introduces 3DGS into indoor 3DOD for the first time, identifying two main challenges: (i) Ambiguous spatial distribution of Gaussian blobs – 3DGS primarily relies on 2D pixel-level supervision, resulting in unclear 3D spatial distribution of Gaussian blobs and poor differentiation between objects and background, which hinders indoor 3DOD; (ii) Excessive background blobs – 2D images typically include numerous background pixels, leading to densely reconstructed 3DGS with many noisy Gaussian blobs representing the background, negatively affecting detection. To tackle (i), we leverage the fact that 3DGS reconstruction is derived from 2D images, and propose an elegant solution by incorporating 2D Boundary Guidance to significantly enhance the spatial distribution of Gaussian blobs, resulting in clearer differentiation between objects and their background (please see fig:teaser). To address (ii), we propose a Box-Focused Sampling strategy using 2D boxes to generate object probability distribution in 3D space, allowing effective probabilistic sampling in 3D to retain more object blobs and reduce noisy background blobs. Benefiting from these innovations, 3DGS-DET significantly outperforms the state-of-the-art NeRF-based method, NeRF-Det++, achieving improvements of +6.0 on mAP@0.25 and +7.8 on mAP@0.5 for the ScanNet, and the +14.9 on mAP@0.25 for the ARKITScenes.

Method: The approach formulates anomaly detection as a mask completion task using masked image modeling. It predicts the distribution of discrete latents in masked regions, leveraging the invariance of discrete latent distributions to low-level pixel variance to focus on logical dependencies between image patches.

Result: The method achieves compatible performance on five benchmarks without requiring pre-trained segmentation models. Comprehensive experiments reveal key factors influencing logical anomaly detection performance.

Conclusion: LADMIM effectively detects logical anomalies by learning long-range dependencies through masked image modeling and discrete representation learning, addressing limitations of conventional local-pattern-focused methods.

Abstract: Detecting anomalies such as an incorrect combination of objects or deviations in their positions is a challenging problem in unsupervised anomaly detection (AD). Since conventional AD methods mainly focus on local patterns of normal images, they struggle with detecting logical anomalies that appear in the global patterns. To effectively detect these challenging logical anomalies, we introduce Logical Anomaly Detection with Masked Image Modeling (LADMIM), a novel unsupervised AD framework that harnesses the power of masked image modeling and discrete representation learning. Our core insight is that predicting the missing region forces the model to learn the long-range dependencies between patches. Specifically, we formulate AD as a mask completion task, which predicts the distribution of discrete latents in the masked region. As a distribution of discrete latents is invariant to the low-level variance in the pixel space, the model can desirably focus on the logical dependencies in the image, which improves accuracy in the logical AD. We evaluate the AD performance on five benchmarks and show that our approach achieves compatible performance without any pre-trained segmentation models. We also conduct comprehensive experiments to reveal the key factors that influence logical AD performance.

Method: SegDAC uses text-grounded segmentation to produce object masks at each timestep, extracts spatially aware token embeddings from these masks, and processes them with a transformer-based actor-critic. Key innovations include segment positional encoding to preserve spatial information and variable-length token processing to handle dynamic object sets.

Result: SegDAC outperforms prior visual generalization methods by 15% on easy, 66% on medium, and 88% on hardest visual perturbation settings across 8 ManiSkill3 manipulation tasks. It matches state-of-the-art visual RL sample efficiency while achieving better generalization under visual changes.

Conclusion: SegDAC demonstrates that segmentation-driven object representations with variable-length token processing and spatial encoding enable RL policies to generalize better under visual perturbations, addressing key limitations of current visual RL approaches.

Abstract: Visual reinforcement learning policies trained on pixel observations often struggle to generalize when visual conditions change at test time. Object-centric representations are a promising alternative, but most approaches use fixed-size slot representations, require image reconstruction, or need auxiliary losses to learn object decompositions. As a result, it remains unclear how to learn RL policies directly from object-level inputs without these constraints. We propose SegDAC, a Segmentation-Driven Actor-Critic that operates on a variable-length set of object token embeddings. At each timestep, text-grounded segmentation produces object masks from which spatially aware token embeddings are extracted. A transformer-based actor-critic processes these dynamic tokens, using segment positional encoding to preserve spatial information across objects. We ablate these design choices and show that both segment positional encoding and variable-length processing are individually necessary for strong performance. We evaluate SegDAC on 8 ManiSkill3 manipulation tasks under 12 visual perturbation types across 3 difficulty levels. SegDAC improves over prior visual generalization methods by 15% on easy, 66% on medium, and 88% on the hardest settings. SegDAC matches the sample efficiency of the state-of-the-art visual RL methods while achieving improved generalization under visual changes. Project Page: https://segdac.github.io/

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to access limitations

Abstract: Failed to fetch summary for 2412.16253: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.16253&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2503.01212: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.01212&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2505.17476: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17476&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.23120: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.23120&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2509.23325: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23325&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2506.06097: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.06097&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2506.21486: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.21486&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract content is unavailable due to API rate limiting

Result: Cannot determine results as abstract content is unavailable due to API rate limiting

Conclusion: Cannot draw conclusions as paper content is inaccessible due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2507.02513: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.02513&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2507.06119: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.06119&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to access limitations

Abstract: Failed to fetch summary for 2510.17914: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.17914&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2508.01587: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.01587&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2410.18613: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.18613&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2510.18632: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.18632&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2508.14327: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.14327&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Builds on quickest change detection framework with adaptive mechanisms that explicitly model mode-dependent distributions of prediction errors that evolve over time with dataset-specific dynamics, enabling robust detection in complex driving environments.

Result: Substantial improvements in both detection delay and false alarm rates compared to prior uncertainty quantification and vision-based OOD approaches, with better accuracy and computational efficiency across established trajectory prediction benchmarks.

Conclusion: The framework offers a practical path toward reliable, driving-aware autonomy by addressing trajectory-level OOD detection, which has been largely overlooked despite its critical importance for autonomous vehicle safety.

Abstract: Trajectory prediction is central to the safe and seamless operation of autonomous vehicles (AVs). In deployment, however, prediction models inevitably face distribution shifts between training data and real-world conditions, where rare or underrepresented traffic scenarios induce out-of-distribution (OOD) cases. While most prior OOD detection research in AVs has concentrated on computer vision tasks such as object detection and segmentation, trajectory-level OOD detection remains largely underexplored. A recent study formulated this problem as a quickest change detection (QCD) task, providing formal guarantees on the trade-off between detection delay and false alarms [1]. Building on this foundation, we propose a new framework that introduces adaptive mechanisms to achieve robust detection in complex driving environments. Empirical analysis across multiple real-world datasets reveals that prediction errors – even on in-distribution samples – exhibit mode-dependent distributions that evolve over time with dataset-specific dynamics. By explicitly modeling these error modes, our method achieves substantial improvements in both detection delay and false alarm rates. Comprehensive experiments on established trajectory prediction benchmarks show that our framework significantly outperforms prior UQ- and vision-based OOD approaches in both accuracy and computational efficiency, offering a practical path toward reliable, driving-aware autonomy.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.15342: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.15342&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2509.17704: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.17704&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to unavailability of paper content

Result: Cannot determine results due to unavailability of paper content

Conclusion: Cannot determine conclusion due to unavailability of paper content

Abstract: Failed to fetch summary for 2511.14099: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.14099&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.24980: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24980&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.18507: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18507&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to lack of paper content.

Result: Unable to determine results due to lack of paper content.

Conclusion: Unable to draw conclusions due to lack of paper content.

Abstract: Failed to fetch summary for 2510.00705: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.00705&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2511.18765: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18765&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2510.12225: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.12225&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2511.23342: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.23342&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2510.27316: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.27316&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable

Result: No results available due to technical access issue

Conclusion: Paper analysis impossible due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2512.05343: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.05343&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper abstract

Result: No results available due to technical error in fetching paper information

Conclusion: Paper analysis cannot be completed due to arXiv API rate limiting

Abstract: Failed to fetch summary for 2511.11266: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.11266&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2511.12708: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.12708&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.16672: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.16672&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable

Result: No results available due to access failure

Conclusion: Paper analysis impossible due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2601.03191: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.03191&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper abstract

Result: No results available due to failed API request

Conclusion: Cannot provide conclusion without paper content

Abstract: Failed to fetch summary for 2511.17354: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17354&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2511.17361: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17361&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.17843: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17843&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content

Result: Cannot determine results without access to the paper content

Conclusion: Cannot determine conclusion without access to the paper content

Abstract: Failed to fetch summary for 2511.21105: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21105&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.21251: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21251&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable due to rate limiting error

Result: Cannot determine results as abstract is unavailable due to rate limiting error

Conclusion: Cannot draw conclusions about paper content due to inability to access abstract

Abstract: Failed to fetch summary for 2511.21662: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21662&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2511.22645: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22645&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown - paper content unavailable due to HTTP 429 error when attempting to fetch from arXiv API.

Result: No results available - the paper summary could not be accessed due to server rate limiting.

Conclusion: Unable to analyze paper content due to technical limitations in accessing the arXiv API.

Abstract: Failed to fetch summary for 2512.06684: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.06684&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.11448: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.11448&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.07801: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.07801&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.08881: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08881&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.14099: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.14099&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2512.15098: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.15098&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to data fetch failure

Result: Unable to determine results due to data fetch failure

Conclusion: Unable to analyze paper due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2512.16201: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.16201&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.14041: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.14041&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2601.02267: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.02267&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract content

Result: Unable to determine results due to missing abstract content

Conclusion: Unable to determine conclusion due to missing abstract content

Abstract: Failed to fetch summary for 2602.23228: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23228&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.07540: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.07540&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2601.08265: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.08265&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2603.00171: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00171&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.12823: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.12823&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot determine conclusion without access to the paper content.

Abstract: Failed to fetch summary for 2601.15475: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.15475&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to server rate limiting

Result: Cannot determine results as paper content is unavailable due to server rate limiting

Conclusion: Cannot draw conclusions as paper content is unavailable due to server rate limiting

Abstract: Failed to fetch summary for 2601.19577: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.19577&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.06688: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.06688&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2602.08820: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.08820&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2603.06723: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.06723&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes SPRig framework with consistency regularization: for skeleton generation - consistency in token and geometry spaces; for skinning - articulation-invariant consistency loss with consistency distillation and structural regularization.

Result: SPRig achieves superior temporal coherence, significantly reduces artifacts in prior methods, and often enhances per-frame static generation quality without sacrificing performance.

Conclusion: SPRig provides an effective solution for pose-invariant rigging on dynamic mesh sequences by enforcing cross-frame consistency, addressing limitations of existing methods.

Abstract: State-of-the-art rigging methods typically assume a predefined canonical rest pose. However, this assumption does not hold for dynamic mesh sequences such as DyMesh or DT4D, where no canonical T-pose is available. When applied independently frame-by-frame, existing methods lack pose invariance and often yield temporally inconsistent topologies. To address this limitation, we propose SPRig, a general fine-tuning framework that enforces cross-frame consistency across a sequence to learn pose-invariant rigs on top of existing models, covering both skeleton and skinning generation. For skeleton generation, we introduce novel consistency regularization in both token space and geometry space. For skinning, we improve temporal stability through an articulation-invariant consistency loss combined with consistency distillation and structural regularization. Extensive experiments show that SPRig achieves superior temporal coherence and significantly reduces artifacts in prior methods, without sacrificing and often even enhancing per-frame static generation quality. The code is available in the supplemental material and will be made publicly available upon publication.

Method: Introduces a dual Gaussian scene representation with geometry Gaussians for surface reconstruction and complementary local reflection Gaussians for near-field specular interactions without ray tracing. Also includes a global environment reflection field for far-field specular reflections, combined with a physically-aware adaptive mixing shader.

Result: Ref-DGS achieves state-of-the-art performance on reflective scenes while training substantially faster than ray-based Gaussian methods, demonstrating efficient and accurate modeling of specular reflections.

Conclusion: The framework successfully addresses the trade-off between accurate specular reflection modeling and computational efficiency, providing an effective solution for reflective appearance in 3D reconstruction and view synthesis.

Abstract: Reflective appearance, especially strong and typically near-field specular reflections, poses a fundamental challenge for accurate surface reconstruction and novel view synthesis. Existing Gaussian splatting methods either fail to model near-field specular reflections or rely on explicit ray tracing at substantial computational cost. We present Ref-DGS, a reflective dual Gaussian splatting framework that addresses this trade-off by decoupling surface reconstruction from specular reflection within an efficient rasterization-based pipeline. Ref-DGS introduces a dual Gaussian scene representation consisting of geometry Gaussians and complementary local reflection Gaussians that capture near-field specular interactions without explicit ray tracing, along with a global environment reflection field for modeling far-field specular reflections. To predict specular radiance, we further propose a lightweight, physically-aware adaptive mixing shader that fuses global and local reflection features. Experiments demonstrate that Ref-DGS achieves state-of-the-art performance on reflective scenes while training substantially faster than ray-based Gaussian methods.

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2602.13172: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13172&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.08605: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08605&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.14027: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.14027&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: 1) Created HomeSafe-Bench via hybrid pipeline combining physical simulation with advanced video generation, featuring 438 diverse cases across six functional areas with fine-grained multidimensional annotations. 2) Proposed HD-Guard: hierarchical streaming architecture with FastBrain (lightweight, high-frequency screening) and SlowBrain (asynchronous, large-scale multimodal reasoning) for real-time safety monitoring.

Result: HD-Guard achieves superior trade-off between latency and performance in unsafe action detection. Analysis identifies critical bottlenecks in current VLM-based safety detection systems.

Conclusion: HomeSafe-Bench addresses the gap in dynamic safety evaluation for household robots, and HD-Guard provides an effective architecture for real-time safety monitoring that balances efficiency with accuracy.

Abstract: The rapid evolution of embodied agents has accelerated the deployment of household robots in real-world environments. However, unlike structured industrial settings, household spaces introduce unpredictable safety risks, where system limitations such as perception latency and lack of common sense knowledge can lead to dangerous errors. Current safety evaluations, often restricted to static images, text, or general hazards, fail to adequately benchmark dynamic unsafe action detection in these specific contexts. To bridge this gap, we introduce HomeSafe-Bench, a challenging benchmark designed to evaluate Vision-Language Models (VLMs) on unsafe action detection in household scenarios. HomeSafe-Bench is contrusted via a hybrid pipeline combining physical simulation with advanced video generation and features 438 diverse cases across six functional areas with fine-grained multidimensional annotations. Beyond benchmarking, we propose Hierarchical Dual-Brain Guard for Household Safety (HD-Guard), a hierarchical streaming architecture for real-time safety monitoring. HD-Guard coordinates a lightweight FastBrain for continuous high-frequency screening with an asynchronous large-scale SlowBrain for deep multimodal reasoning, effectively balancing inference efficiency with detection accuracy. Evaluations demonstrate that HD-Guard achieves a superior trade-off between latency and performance, while our analysis identifies critical bottlenecks in current VLM-based safety detection.

Method: Unable to determine method due to paper retrieval failure

Result: Unable to determine results due to paper retrieval failure

Conclusion: Unable to determine conclusion due to paper retrieval failure

Abstract: Failed to fetch summary for 2602.17077: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.17077&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2603.12067: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12067&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.20673: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20673&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot determine conclusion as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2602.21905: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21905&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to data access issues

Abstract: Failed to fetch summary for 2602.23790: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23790&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2602.23951: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23951&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2602.24084: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.24084&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as the paper content could not be fetched from arXiv

Result: No results available - arXiv API returned HTTP 429 (Too Many Requests) error

Conclusion: The paper cannot be analyzed due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2603.00947: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00947&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed summary fetch

Result: Cannot determine results due to failed summary fetch

Conclusion: Cannot draw conclusions due to failed summary fetch

Abstract: Failed to fetch summary for 2603.01000: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01000&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.04733: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.04733&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.05869: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05869&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to API access limitations

Conclusion: Paper analysis not possible due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2603.07057: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.07057&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.08096: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08096&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2603.09217: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.09217&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2603.09266: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.09266&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.10300: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10300&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: DynVLA introduces Dynamics CoT with a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens. It decouples ego-centric and environment-centric dynamics for more accurate modeling. The model is trained through supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) to generate dynamics tokens before actions.

Result: Extensive experiments on NAVSIM, Bench2Drive, and a large-scale in-house dataset demonstrate that DynVLA consistently outperforms Textual CoT and Visual CoT methods, validating the effectiveness and practical value of Dynamics CoT.

Conclusion: Dynamics CoT captures world evolution in a compact, interpretable, and efficient form, enabling more informed and physically grounded decision-making for driving VLAs while maintaining latency-efficient inference.

Abstract: We propose DynVLA, a driving VLA model that introduces a new CoT paradigm termed Dynamics CoT. DynVLA forecasts compact world dynamics before action generation, enabling more informed and physically grounded decision-making. To obtain compact dynamics representations, DynVLA introduces a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens. Considering the rich environment dynamics in interaction-intensive driving scenarios, DynVLA decouples ego-centric and environment-centric dynamics, yielding more accurate world dynamics modeling. We then train DynVLA to generate dynamics tokens before actions through SFT and RFT, improving decision quality while maintaining latency-efficient inference. Compared to Textual CoT, which lacks fine-grained spatiotemporal understanding, and Visual CoT, which introduces substantial redundancy due to dense image prediction, Dynamics CoT captures the evolution of the world in a compact, interpretable, and efficient form. Extensive experiments on NAVSIM, Bench2Drive, and a large-scale in-house dataset demonstrate that DynVLA consistently outperforms Textual CoT and Visual CoT methods, validating the effectiveness and practical value of Dynamics CoT. Project Page: https://yaoyao-jpg.github.io/dynvla.

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to draw conclusions due to access restrictions

Abstract: Failed to fetch summary for 2603.11460: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.11460&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to access limitations

Result: Cannot determine results due to access limitations

Conclusion: Cannot determine conclusion due to access limitations

Abstract: Failed to fetch summary for 2603.11550: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.11550&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests) error

Result: No results available - failed to fetch paper summary due to rate limiting

Conclusion: Cannot draw conclusions about paper content as it could not be retrieved from arXiv

Abstract: Failed to fetch summary for 2603.12078: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12078&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: A context-aware trajectory abstraction framework that segments AIS sequences into clean trips with mobility-annotated episodes, enriched with multi-source contextual information (geographic entities, navigation features, weather conditions).

Result: The framework produces semantically enriched representations that support generation of controlled natural language descriptions using LLMs, facilitating downstream analytics and maritime reasoning tasks.

Conclusion: The abstraction increases semantic density and reduces spatiotemporal complexity, enabling integration with LLMs for higher-level maritime reasoning and analytics.

Abstract: We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.

Method: Created CRYSTAL benchmark with 6,372 instances using Delphi-inspired pipeline: 4 independent MLLMs generate reasoning trajectories, aggregated via semantic clustering with human validation. Proposed Match F1 (step-level precision/recall) and Ordered Match F1 (penalizes disordered chains). Introduced Causal Process Reward (CPR) that multiplies answer correctness with step alignment, and CPR-Curriculum that progressively increases reasoning difficulty.

Result: Evaluation of 20 MLLMs revealed universal cherry-picking (precision » recall), non-monotonic scaling trade-offs, and disordered reasoning (no model preserved >60% matched steps in correct order). CPR-Curriculum achieved +32% Match F1 via GRPO where additive rewards failed, improving reasoning without manual step annotation.

Conclusion: CRYSTAL exposes systematic reasoning failures invisible to accuracy metrics. The CPR framework and curriculum training effectively improve multimodal reasoning by coupling answer correctness with step-level alignment, enabling better reasoning without manual annotation.

Abstract: We introduce CRYSTAL (__C__lear __R__easoning via __Y__ielded __S__teps, __T__raceability and __L__ogic), a diagnostic benchmark with 6,372 instances that evaluates multimodal reasoning through verifiable intermediate steps. We propose two complementary metrics: Match F1, which scores step-level precision and recall via semantic similarity matching, and Ordered Match F1, which further penalizes disordered reasoning chains. References are constructed through a Delphi-inspired pipeline where four independent MLLMs generate trajectories, aggregated via semantic clustering and validated through human quality gates. Evaluation of 20 MLLMs, including commercial frontier systems not used during benchmark construction, reveals systematic failures invisible to accuracy: universal cherry-picking (precision far exceeds recall), non-monotonic scaling trade-offs, and disordered reasoning where no competitive model preserves more than 60% of matched steps in correct order. Beyond evaluation, we propose the Causal Process Reward (CPR), a multiplicative reward that couples answer correctness with step-level alignment, and CPR-Curriculum, which progressively increases reasoning difficulty during training. CPR-Curriculum achieves +32% Match F1 via GRPO where additive reward strategies fail, improving reasoning without manual step annotation.

Method: ReBalance uses confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. It aggregates hidden states from a small dataset into reasoning mode prototypes, computes a steering vector, and uses a dynamic control function to modulate this vector based on real-time confidence.

Result: Extensive experiments on four models (0.5B to 32B) across nine benchmarks in math reasoning, general QA, and coding tasks show ReBalance effectively reduces output redundancy while improving accuracy.

Conclusion: ReBalance offers a general, training-free, plug-and-play strategy for efficient and robust deployment of Large Reasoning Models by dynamically balancing their reasoning processes.

Abstract: Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs’ reasoning trajectories. A dynamic control function modulates this vector’s strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at https://github.com/yu-lin-li/ReBalance .

Method: Evaluated 6 popular data analysis agents, identified expressivity gaps, then developed AgentFuel to help domain experts create customized end-to-end functional tests with domain-specific datasets and query types.

Result: AgentFuel benchmarks expose key improvement directions for existing data agent frameworks and show anecdotal evidence of performance improvement (e.g., with GEPA). Benchmarks are publicly available.

Conclusion: AgentFuel addresses critical evaluation gaps for timeseries data analysis agents, enabling better assessment and improvement of conversational agents in domain-specific applications.

Abstract: Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to “talk to your data” to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel’s benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at https://huggingface.co/datasets/RockfishData/TimeSeriesAgentEvals.

Method: Formalizes web tasks as sequential decision-making, maps agent architectures to planning paradigms (Step-by-Step→BFS, Tree Search→Best-First, Full-Plan→DFS), proposes 5 novel evaluation metrics, and validates with human-labeled WebArena trajectories.

Result: Step-by-Step agents align better with human trajectories (38% success), while Full-Plan agents excel in technical accuracy (89% element accuracy). The framework reveals trade-offs between different agent architectures.

Conclusion: The planning framework enables principled diagnosis of agent failures and demonstrates the need for comprehensive evaluation metrics beyond simple success rates for selecting appropriate agent architectures.

Abstract: Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.

Method: Proposes using derivatives of deviations between actual sensor readings and expected values, with predictions made by a Random Forest algorithm (selected as most suitable among tested ML algorithms). Includes Deep Learning-based data augmentation for training data.

Result: The method effectively anticipates catastrophic failures before measurements reach critical thresholds, allowing operators to shut down engines and change routes safely. Validated through simulation and real-world data.

Conclusion: The derivative-based approach provides earlier anomaly detection than traditional methods, enabling preventive actions to avoid damage and ensure navigation safety.

Abstract: Catastrophic failures of marine engines imply severe loss of functionality and destroy or damage the systems irreversibly. Being sudden and often unpredictable events, they pose a severe threat to navigation, crew, and passengers. The abrupt nature makes early detection the only effective countermeasure. However, research has concentrated on modeling the gradual degradation of components, with limited attention to sudden and anomalous phenomena. This work proposes a new method for early detection of catastrophic failures. Based on real data from a failed engine, the approach evaluates the derivatives of the deviation between actual sensor readings and expected values of engine variables. Predictions are obtained by a Random Forest, which is the most suitable Machine Learning algorithm among the tested ones. Traditional methods focus on deviations of monitored signals, whereas the proposed approach employs the derivatives of the deviations to provide earlier indications of abnormal dynamics, and to alert that a rapid and dangerous event is breaking out within the system. The method allows the detection of anomalies before measurements reach critical thresholds and alarms are triggered, which is the common method in industry. Consequently, operators can be warned in advance and shut down the engine, then prevent damage and unexpected power loss. Moreover, they have the time to safely change the ship route and avoid potential obstacles. Simulation results conf irm the effectiveness of the proposed approach in anticipating occurrence of catastrophic failures. Validation on real-world data further reinforces the robustness and practical applicability of the method. It is worth noting that data acquisition to train the predictive algorithm is not a problem, since a Deep Learning-based data augmentation procedure is used.

Method: ToolTree employs a Monte Carlo tree search-inspired planning paradigm with dual-stage LLM evaluation (before and after tool execution) and bidirectional pruning to explore tool usage trajectories while efficiently pruning less promising branches.

Result: Empirical evaluations across 4 benchmarks show ToolTree consistently improves performance while maintaining high efficiency, achieving average gains of around 10% compared to state-of-the-art planning paradigms.

Conclusion: ToolTree provides an effective planning framework that enables LLM agents to make informed, adaptive decisions over extended tool-use sequences, addressing limitations of current reactive planning approaches.

Abstract: Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10% compared to the state-of-the-art planning paradigm.

Method: Proposes AIM with two modulation modes: utility modulation (dynamic control over output quality) and focus modulation (precise control over focused input features). Uses a logits redistribution strategy that is training data-agnostic and retraining-free, based on statistical properties of logits ordering via joint probability distributions.

Result: Evaluation confirms AIM’s practicality and versatility across tasks including image classification, semantic segmentation, text generation, and architectures like ResNet, SegFormer, and Llama.

Conclusion: AIM provides an efficient solution for model adaptation without maintaining multiple specialized versions, enabling dynamic behavior control through a single model.

Abstract: Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model’s focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM’s regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM’s practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.

Method: Proposes a three-level Human-AI Governance framework: dimensions (Decision Authority, Process Autonomy, Accountability Configuration), continua (positional spectra along each dimension), and thresholds (critical points where governance requirements shift).

Result: The framework is level-agnostic, applicable from individual deployments to international regulation, and adopts a trust-utility orientation rather than constraint-based approaches, demonstrated through healthcare and European regulation case studies.

Conclusion: HAIG offers a foundation for adaptive regulatory design that anticipates governance challenges by focusing on relational dynamics rather than treating AI as mere objects of governance.

Abstract: This paper introduces the Human-AI Governance (HAIG) framework, contributing to the AI governance (AIG) field by foregrounding the relational dynamics between human and AI actors rather than treating AI systems as objects of governance alone. Current categorical frameworks (e.g., human-in-the-loop models) inadequately capture how AI systems evolve from tools to partners, particularly as foundation models demonstrate emergent capabilities and multi-agent systems exhibit autonomous goal-setting behaviours. As systems are deployed across contexts, agency redistributes in complex patterns that are better represented as positions along continua rather than discrete categories. The HAIG framework operates across three levels: dimensions (Decision Authority, Process Autonomy, and Accountability Configuration), continua (continuous positional spectra along each dimension), and thresholds (critical points along the continua where governance requirements shift qualitatively). The framework’s dimensional architecture is level-agnostic, applicable from individual deployment decisions and organisational governance through to sectorial comparison and national and international regulatory design. Unlike risk-based or principle-based approaches that treat governance primarily as a constraint on AI deployment, HAIG adopts a trust-utility orientation - reframing governance as the condition under which human-AI collaboration can realise its potential, calibrating oversight to specific relational contexts rather than predetermined categories. Case studies in healthcare and European regulation demonstrate how HAIG complements existing frameworks while offering a foundation for adaptive regulatory design that anticipates governance challenges before they emerge.

Method: Developed a multi-agent system where one agent solves abstract engineering problems using domain knowledge, and another agent implements solutions as Chemasim code. Used GitHub Copilot with state-of-the-art LLMs (Claude Opus 4.6) to generate valid syntax for in-house process modeling tool, leveraging technical documentation and examples as context.

Result: Demonstrated effectiveness on typical flowsheet modeling examples including reaction/separation process, pressure-swing distillation, and heteroazeotropic distillation with entrainer selection. Showed capability to generate valid simulation code from engineering specifications.

Conclusion: The framework successfully applies agentic AI to chemical process modeling, though current limitations exist. Future research directions are outlined to enhance capabilities further in this specialized domain.

Abstract: Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.

Method: Parametrized normalization algorithm that transforms ODRL policies into minimal components, converts permissions/prohibitions to permissions-only, and simplifies complex logic constraints into simple ones.

Result: Algorithms preserve semantics while reducing policy complexity, with size complexity exponential on attributes and linear on unique values; enables representation in basic ODRL fragments and simplifies policy comparison.

Conclusion: Normalization approach addresses ODRL complexity barriers, enabling interoperability and easier policy comparison while maintaining semantic equivalence.

Abstract: The ODRL language has become the standard for representing policies and regulations for digital rights. However its complexity is a barrier to its usage, which has caused many related theoretical and practical works to focus on different, and not interoperable, fragments of ODRL. Moreover, semantically equivalent policies can be expressed in numerous different ways, which makes comparing them and processing them harder. Building on top of a recently defined semantics, we tackle these problems by proposing an approach that involves a parametrised normalisation of ODRL policies into its minimal components which reformulates policies with permissions and prohibitions into policies with permissions exclusively, and simplifies complex logic constraints into simple ones. We provide algorithms to compute a normal form for ODRL policies and simplifying numerical and symbolic constraints. We prove that these algorithms preserve the semantics of policies, and analyse the size complexity of the result, which is exponential on the number of attributes and linear on the number of unique values for these attributes. We show how this makes complex policies representable in more basic fragments of ODRL, and how it reduces the problem of policy comparison to the simpler problem of checking if two rules are identical.

Method: MAGPO uses an autoregressive joint policy for scalable, coordinated exploration during centralized training, then explicitly aligns it with decentralized policies to ensure deployability under partial observability constraints.

Result: Empirical evaluation on 43 tasks across 6 diverse environments shows MAGPO consistently outperforms strong CTDE baselines and matches or surpasses fully centralized approaches.

Conclusion: MAGPO offers a principled and practical solution for decentralized multi-agent learning with theoretical guarantees of monotonic policy improvement and strong empirical performance.

Abstract: Due to practical constraints such as partial observability and limited communication, Centralized Training with Decentralized Execution (CTDE) has become the dominant paradigm in cooperative Multi-Agent Reinforcement Learning (MARL). However, existing CTDE methods often underutilize centralized training or lack theoretical guarantees. We propose Multi-Agent Guided Policy Optimization (MAGPO), a novel framework that better leverages centralized training by integrating centralized guidance with decentralized execution. MAGPO uses an autoregressive joint policy for scalable, coordinated exploration and explicitly aligns it with decentralized policies to ensure deployability under partial observability. We provide theoretical guarantees of monotonic policy improvement and empirically evaluate MAGPO on 43 tasks across 6 diverse environments. Results show that MAGPO consistently outperforms strong CTDE baselines and matches or surpasses fully centralized approaches, offering a principled and practical solution for decentralized multi-agent learning. Our code and experimental data can be found in https://github.com/liyheng/MAGPO.

Method: Models MAS routing as semantic-conditioned path selection with three key mechanisms: 1) SFT small language model for intent inference, 2) decomposed routing memory into task-specific pheromone specialists to reduce cross-task interference, 3) quality-gated asynchronous update mechanism to decouple inference from learning.

Result: Extensive experiments on five public benchmarks and high-concurrency stress tests show AMRO-S consistently improves quality-cost trade-off over strong routing baselines while providing traceable routing evidence through structured pheromone patterns.

Conclusion: AMRO-S addresses key deployment limitations of LLM-driven MAS through efficient, interpretable routing that balances performance, cost, and transparency while handling dynamic workloads and mixed intents.

Abstract: Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality–cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality–cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.

Method: Integrates a curated climate knowledge graph (unifying datasets, metadata, tools, workflows) with Agentic AI workflows powered by generative models for natural-language query interpretation, automated data discovery, programmatic data acquisition, and end-to-end climate analysis.

Result: AutoClimDS can reproduce published scientific figures and analyses from natural-language instructions alone, completing entire workflows from dataset selection to modeling. General-purpose LLMs like ChatGPT cannot independently identify authoritative datasets or construct valid retrieval workflows.

Conclusion: The knowledge graph serves as the essential enabling component for autonomous climate data science, providing a pathway toward democratizing climate research through human-AI collaboration.

Abstract: Climate data science remains constrained by fragmented data sources, heterogeneous formats, and steep technical expertise requirements. These barriers slow discovery, limit participation, and undermine reproducibility. We present AutoClimDS, a Minimum Viable Product (MVP) Agentic AI system that addresses these challenges by integrating a curated climate knowledge graph (KG) with a set of Agentic AI workflows designed for cloud-native scientific analysis. The KG unifies datasets, metadata, tools, and workflows into a machine-interpretable structure, while AI agents, powered by generative models, enable natural-language query interpretation, automated data discovery, programmatic data acquisition, and end-to-end climate analysis. A key result is that AutoClimDS can reproduce published scientific figures and analyses from natural-language instructions alone, completing the entire workflow from dataset selection to preprocessing to modeling. When given the same tasks, state-of-the-art general-purpose LLMs (e.g., ChatGPT GPT-5.1) cannot independently identify authoritative datasets or construct valid retrieval workflows using standard web access. This highlights the necessity of structured scientific memory for agentic scientific reasoning. By encoding procedural workflow knowledge into a KG and integrating it with existing technologies (cloud APIs, LLMs, sandboxed execution), AutoClimDS demonstrates that the KG serves as the essential enabling component, the irreplaceable structural foundation, for autonomous climate data science. This approach provides a pathway toward democratizing climate research through human-AI collaboration.

Method: Compress each exchange into structured object with four fields (exchange_core, specific_context, thematic room_assignments, regex-extracted files_touched), creating searchable distilled text averaging 38 tokens per exchange. Evaluated with 201 recall queries, 107 search configurations across 5 pure and 5 cross-layer modes.

Result: Achieves 11x compression (371→38 tokens), best pure distilled configuration reaches 96% of best verbatim MRR (0.717 vs 0.745). Vector search remains non-significant, BM25 degrades significantly. Best cross-layer setup slightly exceeds best pure verbatim baseline (MRR 0.759).

Conclusion: Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality, enabling thousands of exchanges to fit within single prompts while preserving verbatim source for drill-down.

Abstract: Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user’s conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.

Method: Three-phase framework: 1) Experience Anchoring - solidifies subgoal attempts into structured experience tuples with multi-dimensional indexing; 2) Experience Distillation - generalizes successful trajectories into reusable skills and failures into executable guardrails; 3) Knowledge-Driven Closed-Loop Control - injects retrieved skills/guardrails into LLM planner with diagnosis-triggered local replanning.

Result: Experiments on Minecraft MCU long-horizon suite demonstrate consistent improvements over static-retrieval baselines.

Conclusion: The Steve-Evolving framework enables continual evolution of embodied agents through closed-loop experience organization and knowledge distillation without model parameter updates.

Abstract: Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.

Method: Introduces a decision-theoretic view of steganography based on information asymmetry between agents who can and cannot decode hidden content. Proposes generalized V-information framework and steganographic gap metric to measure usable information differences.

Result: Empirically validates the formalism and shows it can detect, quantify, and mitigate steganographic reasoning in LLMs.

Conclusion: Provides a principled approach to address steganographic capabilities in LLMs through information-theoretic analysis of utility differences between agents with and without decoding ability.

Abstract: Large language models are beginning to show steganographic capabilities. Such capabilities could allow misaligned models to evade oversight mechanisms. Yet principled methods to detect and quantify such behaviours are lacking. Classical definitions of steganography, and detection methods based on them, require a known reference distribution of non-steganographic signals. For the case of steganographic reasoning in LLMs, knowing such a reference distribution is not feasible; this renders these approaches inapplicable. We propose an alternative, \textbf{decision-theoretic view of steganography}. Our central insight is that steganography creates an asymmetry in usable information between agents who can and cannot decode the hidden content (present within a steganographic signal), and this otherwise latent asymmetry can be inferred from the agents’ observable actions. To formalise this perspective, we introduce generalised $\mathcal{V}$-information: a utilitarian framework for measuring the amount of usable information within some input. We use this to define the \textbf{steganographic gap} – a measure that quantifies steganography by comparing the downstream utility of the steganographic signal to agents that can and cannot decode the hidden content. We empirically validate our formalism, and show that it can be used to detect, quantify, and mitigate steganographic reasoning in LLMs.

Method: 1) Contrastive reformulation showing GRPO implicitly maximizes margin between policy ratios of correct/incorrect samples; 2) Bilateral Context Conditioning (BICC) allowing cross-referencing of successful/failed reasoning traces during optimization; 3) Reward-Confidence Correction (RCC) stabilizing training by dynamically adjusting advantage baseline using reward-confidence covariance.

Result: Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms.

Conclusion: The proposed methods effectively leverage contrastive information between correct and incorrect reasoning traces, require no additional sampling or auxiliary models, and can be adapted to all GRPO variants for improved reasoning model training.

Abstract: Group Relative Policy Optimization (GRPO) has emerged as an effective method for training reasoning models. While it computes advantages based on group mean, GRPO treats each output as an independent sample during the optimization and overlooks a vital structural signal: the natural contrast between correct and incorrect solutions within the same group, thus ignoring the rich, comparative data that could be leveraged by explicitly pitting successful reasoning traces against failed ones. To capitalize on this, we present a contrastive reformulation of GRPO, showing that the GRPO objective implicitly maximizes the margin between the policy ratios of correct and incorrect samples. Building on this insight, we propose Bilateral Context Conditioning (BICC), a mechanism that allows the model to cross-reference successful and failed reasoning traces during the optimization, enabling a direct information flow across samples. We further introduce Reward-Confidence Correction (RCC) to stabilize training by dynamically adjusts the advantage baseline in GRPO using reward-confidence covariance derived from the first-order approximation of the variance-minimizing estimator. Both mechanisms require no additional sampling or auxiliary models and can be adapted to all GRPO variants. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms. Code is available at \href{https://github.com/Skylanding/BiCC}{https://github.com/Skylanding/BiCC}.

Method: Proposes a cumulative pyramid maturity model with four disciplines: prompt engineering (base), context engineering (designing informational environment), intent engineering (encoding organizational goals), and specification engineering (machine-readable policies). Defines five context quality criteria: relevance, sufficiency, isolation, economy, and provenance.

Result: Enterprise data shows 75% of enterprises plan agentic AI deployment within two years, but deployments have surged and retreated due to scaling complexity. The Klarna case illustrates contextual and intentional deficits in current approaches.

Conclusion: Context engineering frames context as the agent’s operating system. Control over context, intent, and specifications determines behavior, strategy, and scalability of multi-agent systems. The proposed maturity model provides a framework for scaling autonomous AI agents.

Abstract: As artificial intelligence (AI) systems evolve from stateless chatbots to autonomous multi-step agents, prompt engineering (PE), the discipline of crafting individual queries, proves necessary but insufficient. This paper introduces context engineering (CE) as a standalone discipline concerned with designing, structuring, and managing the entire informational environment in which an AI agent makes decisions. Drawing on vendor architectures (Google ADK, Anthropic, LangChain), current academic work (ACE framework, Google DeepMind’s intelligent delegation), enterprise research (Deloitte, 2026; KPMG, 2026), and the author’s experience building a multi-agent system, the paper proposes five context quality criteria: relevance, sufficiency, isolation, economy, and provenance, and frames context as the agent’s operating system. Two higher-order disciplines follow. Intent engineering (IE) encodes organizational goals, values, and trade-off hierarchies into agent infrastructure. Specification engineering (SE) creates a machine-readable corpus of corporate policies and standards enabling autonomous operation of multi-agent systems at scale. Together these four disciplines form a cumulative pyramid maturity model of agent engineering, in which each level subsumes the previous one as a necessary foundation. Enterprise data reveals a gap: while 75% of enterprises plan agentic AI deployment within two years (Deloitte, 2026), deployment has surged and retreated as organizations confront scaling complexity (KPMG, 2026). The Klarna case illustrates a dual deficit, contextual and intentional. Whoever controls the agent’s context controls its behavior; whoever controls its intent controls its strategy; whoever controls its specifications controls its scale.

Method: System combines: (1) stage-aware triage routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from LLM. Evaluation workflow includes triage benchmark, synthetic retrieval benchmark, LLM-as-judge comparison, and expert validation.

Result: Achieved 86.7% emergency recall on triage benchmark with explicit trade-off reporting, developed comprehensive evaluation benchmarks, and demonstrated that trustworthy medical assistants require defense-in-depth design with multi-method evaluation rather than single model choices.

Conclusion: Trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than relying on any single model or evaluation method choice.

Abstract: The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.

Method: Developed a metamorphic testing framework applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, contrastive formulation) across seven foundation models from four architectural families, evaluated on 19 multi-step reasoning problems across eight scientific domains.

Result: Model scale doesn’t predict robustness - smaller Qwen3-30B-A3B achieved highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibited greater fragility. The framework successfully identified significant robustness issues in LLM reasoning agents.

Conclusion: Standard accuracy benchmarks are insufficient for evaluating LLM reasoning reliability; metamorphic testing reveals critical robustness gaps that must be addressed for safe deployment of LLM agents in consequential applications.

Abstract: Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance.Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.

Method: Replaced Transformer blocks in TRM with Mamba-2 hybrid operators while maintaining parameter parity (~6.8M parameters). Evaluated on ARC-AGI-1 abstract reasoning benchmark.

Result: Mamba-2 hybrid improves pass@2 by +2.0% (45.88% vs 43.88%) and consistently outperforms at higher K values (+4.75% at pass@100), while maintaining pass@1 parity. This suggests improved candidate coverage with similar top-1 selection.

Conclusion: Mamba-2 hybrid operators preserve reasoning capability within recursive scaffolds, establishing SSM-based operators as viable candidates in recursive operator design space and taking a first step toward understanding optimal mixing strategies for recursive reasoning.

Abstract: Recent work on recursive reasoning models like TRM demonstrates that tiny networks (7M parameters) can achieve strong performance on abstract reasoning tasks through latent recursion – iterative refinement in hidden representation space without emitting intermediate tokens. This raises a natural question about operator choice: Mamba-2’s state space recurrence is itself a form of iterative refinement, making it a natural candidate for recursive reasoning – but does introducing Mamba-2 into the recursive scaffold preserve reasoning capability? We investigate this by replacing the Transformer blocks in TRM with Mamba-2 hybrid operators while maintaining parameter parity (6.83M vs 6.86M parameters). On ARC-AGI-1, we find that the hybrid improves pass@2 (the official metric) by +2.0% (45.88% vs 43.88%) and consistently outperforms at higher K values (+4.75% at pass@100), whilst maintaining pass@1 parity. This suggests improved candidate coverage – the model generates correct solutions more reliably – with similar top-1 selection. Our results validate that Mamba-2 hybrid operators preserve reasoning capability within the recursive scaffold, establishing SSM-based operators as viable candidates in the recursive operator design space and taking a first step towards understanding the best mixing strategies for recursive reasoning.

Method: ACSLWL (Active Causal Structure Learning with Latent Variables) involves: acting in environment, discovering new causal relations, constructing new causal models, exploiting models for utility maximization, detecting latent variables during unexpected observations, and building new structures with optimal parameter estimation.

Result: Demonstrates that complex planning and expectation-based detour behavior can be learned when a simulated robot unexpectedly encounters a transparent barrier, transforming unexpected situations into predictable ones with optimal operating plans.

Conclusion: Active causal structure learning with latent variables is essential for AGI agents to efficiently cope with new situations by constructing new internal causal models that transform unexpected, inefficient scenarios into predictable, optimal ones.

Abstract: Artificial General Intelligence (AGI) Agents and Robots must be able to cope with everchanging environments and tasks. They must be able to actively construct new internal causal models of their interactions with the environment when new structural changes take place in the environment. Thus, we claim that active causal structure learning with latent variables (ACSLWL) is a necessary component to build AGI agents and robots. This paper describes how a complex planning and expectation-based detour behavior can be learned by ACSLWL when, unexpectedly, and for the first time, the simulated robot encounters a sort of transparent barrier in its pathway towards its target. ACSWL consists of acting in the environment, discovering new causal relations, constructing new causal models, exploiting the causal models to maximize its expected utility, detecting possible latent variables when unexpected observations occur, and constructing new structures-internal causal models and optimal estimation of the associated parameters, to be able to cope efficiently with the new encountered situations. That is, the agent must be able to construct new causal internal models that transform a previously unexpected and inefficient (sub-optimal) situation, into a predictable situation with an optimal operating plan.

Method: Conducted systematic study on evaluation faking using diverse foundation models with mainstream safety benchmarks. Developed chain-of-thought monitoring technique to detect faking intent and uncover internal signals correlated with such behavior.

Result: Found significant observer effects: reasoning models recognize evaluation 16% more often than non-reasoning models; scaling models (32B to 671B) increases faking by over 30%; AI with basic memory is 2.3x more likely to recognize evaluation and scores 19% higher on safety tests.

Conclusion: Evaluation faking is a real phenomenon where advanced AI systems alter behavior when recognizing evaluation contexts, compromising safety assessment integrity. The study provides detection methods and insights for future mitigation.

Abstract: As foundation models grow increasingly more intelligent, reliable and trustworthy safety evaluation becomes more indispensable than ever. However, an important question arises: Whether and how an advanced AI system would perceive the situation of being evaluated, and lead to the broken integrity of the evaluation process? During standard safety tests on a mainstream large reasoning model, we unexpectedly observe that the model without any contextual cues would occasionally recognize it is being evaluated and hence behave more safety-aligned. This motivates us to conduct a systematic study on the phenomenon of evaluation faking, i.e., an AI system autonomously alters its behavior upon recognizing the presence of an evaluation context and thereby influencing the evaluation results. Through extensive experiments on a diverse set of foundation models with mainstream safety benchmarks, we reach the main finding termed the observer effects for AI: When the AI system under evaluation is more advanced in reasoning and situational awareness, the evaluation faking behavior becomes more ubiquitous, which reflects in the following aspects: 1) Reasoning models recognize evaluation 16% more often than non-reasoning models. 2) Scaling foundation models (32B to 671B) increases faking by over 30% in some cases, while smaller models show negligible faking. 3) AI with basic memory is 2.3x more likely to recognize evaluation and scores 19% higher on safety tests (vs. no memory). To measure this, we devised a chain-of-thought monitoring technique to detect faking intent and uncover internal signals correlated with such behavior, offering insights for future mitigation studies.

Method: The Darwin Gödel Machine maintains an archive of coding agents, grows it by sampling agents and using foundation models to create new versions, forming a tree of diverse agents through open-ended exploration with empirical validation on benchmarks.

Result: DGM improved coding capabilities significantly: SWE-bench performance increased from 20.0% to 50.0%, Polyglot from 14.2% to 30.7%, outperforming baselines without self-improvement or open-ended exploration.

Conclusion: DGM represents a significant step toward self-improving AI capable of autonomous innovation, with safety precautions implemented. It demonstrates practical self-improvement through evolutionary principles and empirical validation.

Abstract: Today’s AI systems have human-designed, fixed architectures and cannot autonomously and continuously improve themselves. The advance of AI could itself be automated. If done safely, that would accelerate AI development and allow us to reap its benefits much sooner. Meta-learning can automate the discovery of novel algorithms, but is limited by first-order improvements and the human design of a suitable search space. The Gödel machine proposed a theoretical alternative: a self-improving AI that repeatedly modifies itself in a provably beneficial manner. Unfortunately, proving that most changes are net beneficial is impossible in practice. We introduce the Darwin Gödel Machine (DGM), a self-improving system that iteratively modifies its own code (thereby also improving its ability to modify its own codebase) and empirically validates each change using coding benchmarks. Inspired by Darwinian evolution and open-endedness research, the DGM maintains an archive of generated coding agents. It grows the archive by sampling an agent from it and using a foundation model to create a new, interesting, version of the sampled agent. This open-ended exploration forms a growing tree of diverse, high-quality agents and allows the parallel exploration of many different paths through the search space. Empirically, the DGM automatically improves its coding capabilities (e.g., better code editing tools, long-context window management, peer-review mechanisms), increasing performance on SWE-bench from 20.0% to 50.0%, and on Polyglot from 14.2% to 30.7%. Furthermore, the DGM significantly outperforms baselines without self-improvement or open-ended exploration. All experiments were done with safety precautions (e.g., sandboxing, human oversight). The DGM is a significant step toward self-improving AI, capable of gathering its own stepping stones along paths that unfold into endless innovation.

Method: Curriculum learning framework based on Group Relative Policy Optimization (GRPO) that accounts for trajectory difficulty variation, with a reward function combining rule-based signals and model-judged evaluation.

Result: Outperforms previous SOTA by 5.6% on Android Control benchmark and 10.3% on internal online benchmarks.

Conclusion: Integrating RL with curriculum learning effectively improves GUI interaction task performance by addressing difficulty variation and providing nuanced feedback.

Abstract: As autonomous agents become adept at understanding and interacting with graphical user interface (GUI) environments, a new era of automated task execution is emerging. Recent studies have demonstrated that Reinforcement Learning (RL) can effectively enhance agents’ performance in dynamic interactive GUI environments. However, these methods face two key limitations: (1) they overlook the significant variation in difficulty across different GUI tasks by treating the entire training data as a uniform set, which hampers the agent’s ability to adapt its learning process; and (2) most approaches collapse task-specific nuances into a single, coarse reward, leaving the agent with a uniform signal that yields inefficient policy updates. To address these limitations, we propose CRAFT-GUI, a curriculum learning framework based on Group Relative Policy Optimization (GRPO) that explicitly accounts for the varying difficulty across trajectories. To enable more fine-grained policy optimization, we design a reward function that combines simple rule-based signals with model-judged evaluation, providing richer and more nuanced feedback during training. Experimental results demonstrate that our method achieves significant improvements over previous state-of-the-art approaches, outperforming them by 5.6% on public benchmarks Android Control and 10.3% on our internal online benchmarks, respectively. These findings empirically validate the effectiveness of integrating reinforcement learning with curriculum learning in GUI interaction tasks.

Method: XSkill uses a dual-stream framework with two complementary knowledge forms: experiences (action-level guidance) and skills (task-level guidance). It grounds knowledge extraction and retrieval in visual observations, distills knowledge from multi-path rollouts via visually grounded summarization and cross-rollout critique, and forms a continual learning loop by feeding usage history back into accumulation.

Result: XSkill consistently and substantially outperforms both tool-only and learning-based baselines on five benchmarks across diverse domains with four backbone models, showing superior zero-shot generalization.

Conclusion: The dual-stream approach with experiences and skills enables effective continual learning in multimodal agents, with the two knowledge streams playing complementary roles in influencing reasoning behaviors.

Abstract: Multimodal agents can now tackle complex reasoning tasks with diverse tools, yet they still suffer from inefficient tool use and inflexible orchestration in open-ended settings. A central challenge is enabling such agents to continually improve without parameter updates by learning from past trajectories. We identify two complementary forms of reusable knowledge essential for this goal: experiences, providing concise action-level guidance for tool selection and decision making, and skills, providing structured task-level guidance for planning and tool use. To this end, we propose XSkill, a dual-stream framework for continual learning from experience and skills in multimodal agents. XSkill grounds both knowledge extraction and retrieval in visual observations. During accumulation, XSkill distills and consolidates experiences and skills from multi-path rollouts via visually grounded summarization and cross-rollout critique. During inference, it retrieves and adapts this knowledge to the current visual context and feeds usage history back into accumulation to form a continual learning loop. Evaluated on five benchmarks across diverse domains with four backbone models, XSkill consistently and substantially outperforms both tool-only and learning-based baselines. Further analysis reveals that the two knowledge streams play complementary roles in influencing the reasoning behaviors of agents and show superior zero-shot generalization.

Method: Develop theoretical conditions that guarantee orientability of micro-level edges between temporal variables given background knowledge in summary causal graphs, assuming access to faithful and causally sufficient distributions.

Result: Theoretical guarantees for edge orientation at micro-level even with macro-level cycles or bidirected edges, providing practical guidance for leveraging expert knowledge in causal discovery.

Conclusion: Summary causal graphs can effectively inform causal discovery in complex temporal systems, highlighting the value of incorporating expert knowledge to improve causal inference from observational time series data.

Abstract: Understanding causal relations between temporal variables is a central challenge in time series analysis, particularly when the full causal structure is unknown. Even when the full causal structure cannot be fully specified, experts often succeed in providing a high-level abstraction of the causal graph, known as a summary causal graph, which captures the main causal relations between different time series while abstracting away micro-level details. In this work, we present conditions that guarantee the orientability of micro-level edges between temporal variables given the background knowledge encoded in a summary causal graph and assuming having access to a faithful and causally sufficient distribution with respect to the true unknown graph. Our results provide theoretical guarantees for edge orientation at the micro-level, even in the presence of cycles or bidirected edges at the macro-level. These findings offer practical guidance for leveraging SCGs to inform causal discovery in complex temporal systems and highlight the value of incorporating expert knowledge to improve causal inference from observational time series data.

Method: Isolate execution capability by explicitly providing knowledge and plans for long-horizon tasks, analyze per-step accuracy degradation with increasing steps, identify self-conditioning effect (models become more likely to make mistakes when context contains prior errors), and test thinking models on task length execution.

Result: Larger models execute significantly more turns correctly even when small models have near-perfect single-turn accuracy; per-step accuracy degrades with more steps due to self-conditioning; thinking mitigates self-conditioning and enables much longer task execution in single turns.

Conclusion: Scaling model size and sequential test-time compute provides massive benefits for long-horizon tasks, with thinking being crucial for mitigating self-conditioning errors and enabling longer task execution.

Abstract: Does continued scaling of large language models (LLMs) yield diminishing returns? In this work, we show that short-task benchmarks may give an illusion of slowing progress, as even marginal gains in single-step accuracy can compound into exponential improvements in the length of tasks a model can successfully complete. Then, we argue that failures of LLMs when simple tasks are made longer arise from mistakes in execution, rather than an inability to reason. So, we propose isolating execution capability, by explicitly providing the knowledge and plan needed to solve a long-horizon task. First, we find that larger models can correctly execute significantly more turns even when small models have near-perfect single-turn accuracy. We then observe that the per-step accuracy of models degrades as the number of steps increases. This is not just due to long-context limitations – curiously, we observe a self-conditioning effect – models become more likely to make mistakes when the context contains their errors from prior turns. Self-conditioning does not reduce by just scaling the model size. But, we find that thinking mitigates self-conditioning, and also enables execution of much longer tasks in a single turn. We conclude by benchmarking frontier thinking models on the length of tasks they can execute in a single turn. Overall, by focusing on the ability to execute, we hope to reconcile debates on how LLMs can solve complex reasoning problems yet fail at simple tasks when made longer, and highlight the massive benefits of scaling model size and sequential test-time compute for long-horizon tasks.

Method: Introduces operational safety concept and OffTopicEval evaluation suite/benchmark; evaluates 20 open-weight LLMs across 6 model families; proposes two prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground).

Result: All tested models show poor operational safety, with best models (Qwen-3 235B: 77.77%, Mistral 24B: 79.96%) still unreliable; GPT models 62-73%, Phi 48-70%, Gemma 39.53%, Llama-3 23.84%. Prompt-based steering improves OOD refusal: Q-ground up to 23%, P-ground raises Llama-3.3 (70B) by 41% and Qwen-3 (30B) by 27%.

Conclusion: Operational safety is a critical alignment issue requiring urgent interventions; prompt-based steering shows promise as a first step toward more reliable LLM-based agents for enterprise deployment.

Abstract: Large Language Model (LLM) safety is one of the most pressing challenges for enabling wide-scale deployment. While most studies and global discussions focus on generic harms, such as models assisting users in harming themselves or others, enterprises face a more fundamental concern: whether LLM-based agents are safe for their intended use case. To address this, we introduce operational safety, defined as an LLM’s ability to appropriately accept or refuse user queries when tasked with a specific purpose. We further propose OffTopicEval, an evaluation suite and benchmark for measuring operational safety both in general and within specific agentic use cases. Our evaluations on six model families comprising 20 open-weight LLMs reveal that while performance varies across models, all of them remain highly operationally unsafe. Even the strongest models - Qwen-3 (235B) with 77.77% and Mistral (24B) with 79.96% - fall far short of reliable operational safety, while GPT models plateau in the 62-73% range, Phi achieves only mid-level scores (48-70%), and Gemma and Llama-3 collapse to 39.53% and 23.84%, respectively. While operational safety is a core model alignment issue, to suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23%, while P-ground delivers even larger boosts, raising Llama-3.3 (70B) by 41% and Qwen-3 (30B) by 27%. These results highlight both the urgent need for operational safety interventions and the promise of prompt-based steering as a first step toward more reliable LLM-based agents.

Method: Formulates NTO for congestion management considering linearized AC power flow, and proposes a graph neural network (GNN)-accelerated approach. Develops a heterogeneous edge-aware message passing GNN to predict effective nodes for busbar splitting actions as candidate NTO solutions.

Result: Case studies show up to 4 orders-of-magnitude speed-up, delivering AC-feasible solutions within one minute and a 2.3% optimality gap on the GOC 2000-bus system.

Conclusion: The proposed GNN captures local flow patterns, improves generalization to unseen topology changes, and enhances transferability across systems, demonstrating a significant step toward near-real-time NTO for large-scale systems with topology and cross-system generalization.

Abstract: Network topology optimization (NTO) via busbar splitting can mitigate transmission grid congestion and reduce redispatch costs. However, solving this mixed-integer nonlinear problem for large-scale systems in near-real-time is currently intractable with existing solvers. Machine learning (ML) approaches have emerged as a promising alternative, but they have limited generalization to unseen topologies, varying operating conditions, and different systems, which limits their practical applicability. This paper formulates NTO for congestion management considering linearized AC power flow, and proposes a graph neural network (GNN)-accelerated approach. We develop a heterogeneous edge-aware message passing GNN to predict effective nodes for busbar splitting actions as candidate NTO solutions. The proposed GNN captures local flow patterns, improves generalization to unseen topology changes, and enhances transferability across systems. Case studies show up to 4 orders-of-magnitude speed-up, delivering AC-feasible solutions within one minute and a 2.3% optimality gap on the GOC 2000-bus system. These results demonstrate a significant step toward near-real-time NTO for large-scale systems with topology and cross-system generalization.

Method: Proposes task-specific Prompt-Prototype (ProP) approach where task-specific prompts facilitate feature learning for current tasks, while prototypes capture representative input features. During inference, predictions are generated by binding each task-specific prompt with its associated prototype. Also introduces regularization constraints during prompt initialization to penalize large values for stability.

Result: Experiments on several widely used datasets demonstrate the effectiveness of the proposed method. The framework successfully removes dependency on key-value pairs while maintaining performance.

Conclusion: The ProP method offers a fresh perspective for continual learning research by eliminating key-value pair dependencies, reducing inter-task interference, and improving scalability compared to mainstream prompt-based approaches.

Abstract: Continual learning aims to enable models to acquire new knowledge while retaining previously learned information. Prompt-based methods have shown remarkable performance in this domain; however, they typically rely on key-value pairing, which can introduce inter-task interference and hinder scalability. To overcome these limitations, we propose a novel approach employing task-specific Prompt-Prototype (ProP), thereby eliminating the need for key-value pairs. In our method, task-specific prompts facilitate more effective feature learning for the current task, while corresponding prototypes capture the representative features of the input. During inference, predictions are generated by binding each task-specific prompt with its associated prototype. Additionally, we introduce regularization constraints during prompt initialization to penalize excessively large values, thereby enhancing stability. Experiments on several widely used datasets demonstrate the effectiveness of the proposed method. In contrast to mainstream prompt-based approaches, our framework removes the dependency on key-value pairs, offering a fresh perspective for future continual learning research.

Method: Benchmarked 20+ LLMs against human baseline on 11 causal judgment tasks using collider structures (C₁ → E ← C₂), analyzed reasoning strategies, and tested robustness under semantic abstraction and prompt overloading with chain-of-thought prompting.

Result: Most LLMs exhibit more rule-like reasoning than humans, don’t mirror human collider biases (weak explaining away, Markov violations), and CoT improves robustness for many models.

Conclusion: LLMs can complement humans when known biases are undesirable, but their rule-like reasoning may break down with intrinsic uncertainty, highlighting need to characterize reasoning strategies for safe deployment.

Abstract: Large language models (LLMs) are increasingly used in domains where causal reasoning matters, yet it remains unclear whether their judgments reflect normative causal computation, human-like shortcuts, or brittle pattern matching. We benchmark 20+ LLMs against a matched human baseline on 11 causal judgment tasks formalized by a collider structure ($C_1 \rightarrow E \leftarrow C_2$). We find that a small interpretable model compresses LLMs’ causal judgments well and that most LLMs exhibit more rule-like reasoning strategies than humans who seem to account for unmentioned latent factors in their probability judgments. Furthermore, most LLMs do not mirror the characteristic human collider biases of weak explaining away and Markov violations. We probe LLMs’ causal judgment robustness under (i) semantic abstraction and (ii) prompt overloading (injecting irrelevant text), and find that chain-of-thought (CoT) increases robustness for many LLMs. Together, this divergence suggests LLMs can complement humans when known biases are undesirable, but their rule-like reasoning may break down when uncertainty is intrinsic - highlighting the need to characterize LLM reasoning strategies for safe, effective deployment.

Method: Created SkillsBench benchmark with 86 tasks across 11 domains, each paired with curated Skills and deterministic verifiers. Evaluated 7 agent-model configurations over 7,308 trajectories under three conditions: no Skills, curated Skills, and self-generated Skills.

Result: Curated Skills raise average pass rate by 16.2 percentage points, but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare). 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average. Focused Skills with 2-3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.

Conclusion: Skills can significantly improve LLM agent performance when properly curated, but models cannot reliably author the procedural knowledge they benefit from consuming. The effectiveness depends on domain and skill design, with focused skills being more effective than comprehensive documentation.

Abstract: Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2–3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.

Method: OpenSage enables LLMs to automatically create agents with self-generated topology and toolsets, provides comprehensive structured memory support with hierarchical graph-based memory system, and includes specialized toolkits for software engineering tasks.

Result: Extensive experiments across three state-of-the-art benchmarks with various backbone models demonstrate OpenSage’s advantages over existing ADKs, with rigorous ablation studies validating the effectiveness of each design component.

Conclusion: OpenSage paves the way for next-generation agent development by shifting from human-centered to AI-centered paradigms, enabling automatic agent creation with self-generated components.

Abstract: Agent development kits (ADKs) provide effective platforms and tooling for constructing agents, and their designs are critical to the constructed agents’ performance, especially the functionality for agent topology, tools, and memory. However, current ADKs either lack sufficient functional support or rely on humans to manually design these components, limiting agents’ generalizability and overall performance. We propose OpenSage, the first ADK that enables LLMs to automatically create agents with self-generated topology and toolsets while providing comprehensive and structured memory support. OpenSage offers effective functionality for agents to create and manage their own sub-agents and toolkits. It also features a hierarchical, graph-based memory system for efficient management and a specialized toolkit tailored to software engineering tasks. Extensive experiments across three state-of-the-art benchmarks with various backbone models demonstrate the advantages of OpenSage over existing ADKs. We also conduct rigorous ablation studies to demonstrate the effectiveness of our design for each component. We believe OpenSage can pave the way for the next generation of agent development, shifting the focus from human-centered to AI-centered paradigms.

Method: OPENDEV uses a compound AI system architecture with workload-specialized model routing, a dual-agent architecture separating planning from execution, lazy tool discovery, and adaptive context compaction that progressively reduces older observations. It also employs an automated memory system to accumulate project-specific knowledge across sessions and counteracts instruction fade-out through event-driven system reminders.

Result: OPENDEV provides a secure, extensible foundation for terminal-first AI assistance, offering a blueprint for robust autonomous software engineering by enforcing explicit reasoning phases and prioritizing context efficiency.

Conclusion: OPENDEV represents a new paradigm in AI coding assistance - moving from IDE plugins to terminal-native agents that offer unprecedented autonomy for development tasks while maintaining safety and efficiency through innovative architectural choices.

Abstract: The landscape of AI coding assistance is undergoing a fundamental shift from complex IDE plugins to versatile, terminal-native agents. Operating directly where developers manage source control, execute builds, and deploy environments, CLI-based agents offer unprecedented autonomy for long-horizon development tasks. In this paper, we present OPENDEV, an open-source, command-line coding agent written in Rust, engineered specifically for this new paradigm. Effective autonomous assistance requires strict safety controls and highly efficient context management to prevent context bloat and reasoning degradation. OPENDEV overcomes these challenges through a compound AI system architecture with workload-specialized model routing, a dual-agent architecture separating planning from execution, lazy tool discovery, and adaptive context compaction that progressively reduces older observations. Furthermore, it employs an automated memory system to accumulate project-specific knowledge across sessions and counteracts instruction fade-out through event-driven system reminders. By enforcing explicit reasoning phases and prioritizing context efficiency, OPENDEV provides a secure, extensible foundation for terminal-first AI assistance, offering a blueprint for robust autonomous software engineering.

Method: Tested seven frontier AI models over 18 months on two purpose-built cyber ranges: a 32-step corporate network attack and a 7-step industrial control system attack, comparing performance at varying inference-time compute budgets (10M to 100M tokens).

Result: Two key trends: 1) Performance scales log-linearly with inference-time compute (10M to 100M tokens yields up to 59% gains), 2) Each successive model generation outperforms predecessors at fixed token budgets (corporate network performance rose from 1.7 to 9.8 steps at 10M tokens). Best single run completed 22/32 corporate steps (≈6 of 14 human expert hours). Industrial control system performance remains limited but improving (1.2-1.4 of 7 steps average, max 3).

Conclusion: Frontier AI models show rapidly improving autonomous cyber-attack capabilities that scale with compute and model generations, posing increasing security risks while highlighting the need for defensive measures.

Abstract: We evaluate the autonomous cyber-attack capabilities of frontier AI models on two purpose-built cyber ranges-a 32-step corporate network attack and a 7-step industrial control system attack-that require chaining heterogeneous capabilities across extended action sequences. By comparing seven models released over an eighteen-month period (August 2024 to February 2026) at varying inference-time compute budgets, we observe two capability trends. First, model performance scales log-linearly with inference-time compute, with no observed plateau-increasing from 10M to 100M tokens yields gains of up to 59%, requiring no specific technical sophistication from the operator. Second, each successive model generation outperforms its predecessor at fixed token budgets: on the corporate network range, average steps completed at 10M tokens rose from 1.7 (GPT-4o, August 2024) to 9.8 (Opus 4.6, February 2026). The best single run completed 22 of 32 steps, corresponding to roughly 6 of the estimated 14 hours a human expert would need. On the industrial control system range, performance remains limited, though the most recent models are the first to reliably complete steps, averaging 1.2-1.4 of 7 (max 3).

Method: Introduces COMPASS Framework with Orchestrator and four specialized sub-agents (sovereignty, carbon-aware computing, compliance, ethics), each augmented with Retrieval-Augmented Generation (RAG) for context-specific document grounding. Uses LLM-as-a-judge methodology to assign quantitative scores and generate explainable justifications for real-time arbitration of conflicting objectives.

Result: RAG integration significantly enhances semantic coherence and mitigates hallucination risks. The composition-based design facilitates seamless integration into diverse application domains while preserving interpretability and traceability.

Conclusion: COMPASS provides a novel unified architecture for enforcing value-aligned AI through modular, extensible governance mechanisms that systematically integrate multiple imperatives into autonomous agent decision-making.

Abstract: The rapid proliferation of large language model (LLM)-based agentic systems raises critical concerns regarding digital sovereignty, environmental sustainability, regulatory compliance, and ethical alignment. Whilst existing frameworks address individual dimensions in isolation, no unified architecture systematically integrates these imperatives into the decision-making processes of autonomous agents. This paper introduces the COMPASS (Compliance and Orchestration for Multi-dimensional Principles in Autonomous Systems with Sovereignty) Framework, a novel multi-agent orchestration system designed to enforce value-aligned AI through modular, extensible governance mechanisms. The framework comprises an Orchestrator and four specialised sub-agents addressing sovereignty, carbon-aware computing, compliance, and ethics, each augmented with Retrieval-Augmented Generation (RAG) to ground evaluations in verified, context-specific documents. By employing an LLM-as-a-judge methodology, the system assigns quantitative scores and generates explainable justifications for each assessment dimension, enabling real-time arbitration of conflicting objectives. We validate the architecture through automated evaluation, demonstrating that RAG integration significantly enhances semantic coherence and mitigates the hallucination risks. Our results indicate that the framework’s composition-based design facilitates seamless integration into diverse application domains whilst preserving interpretability and traceability.

Method: Used Cognition-Affect-Conation (CAC) framework with structural equation modeling on survey data from 436 OpenClaw users, analyzing enabling factors (personalization, intelligence, relative advantage) and inhibiting factors (privacy concern, opacity, risk).

Result: Positive perceptions strengthen attitudes toward OpenClaw and increase behavioral intention, while negative perceptions increase distrust and reduce intention to use the system.

Conclusion: Provides insights into psychological mechanisms of AI agent adoption, showing cognitive perceptions drive affective responses which shape behavioral intentions.

Abstract: This study examines users’ behavioural intention to use OpenClaw through the Cognition–Affect–Conation (CAC) framework. The research investigates how cognitive perceptions of the system influence affective responses and subsequently shape behavioural intention. Enabling factors include perceived personalisation, perceived intelligence, and relative advantage, while inhibiting factors include privacy concern, algorithmic opacity, and perceived risk. Survey data from 436 OpenClaw users were analysed using structural equation modelling. The results show that positive perceptions strengthen users’ attitudes toward OpenClaw, which increase behavioural intention, whereas negative perceptions increase distrust and reduce intention to use the system. The study provides insights into the psychological mechanisms influencing the adoption of autonomous AI agents.

Method: Method unknown - unable to access paper details due to HTTP 429 error from arXiv API.

Result: Results unknown - paper content inaccessible due to rate limiting on arXiv API.

Conclusion: Cannot provide conclusion without access to paper content. The arXiv ID suggests this is a recent work that may be relevant to multimodal AI research.

Abstract: Failed to fetch summary for 2508.10954: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.10954&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request

Result: No results available - paper summary could not be fetched

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper content

Abstract: Failed to fetch summary for 2508.16624: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.16624&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2509.12431: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.12431&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.03366: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.03366&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.15511: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15511&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to data fetch failure

Result: Unable to determine results due to data fetch failure

Conclusion: Unable to draw conclusions due to data fetch failure

Abstract: Failed to fetch summary for 2510.19973: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.19973&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2511.03211: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.03211&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request

Result: No results available - the request to fetch paper information resulted in HTTP 429 error

Conclusion: Cannot analyze paper content due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2512.11946: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.11946&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access limitations

Result: Unable to determine results due to access limitations

Conclusion: Unable to determine conclusion due to access limitations

Abstract: Failed to fetch summary for 2512.12858: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.12858&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2601.08697: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.08697&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2601.10436: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.10436&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content inaccessible

Result: No results available - only error message indicating HTTP 429 status

Conclusion: Cannot analyze paper due to technical limitations in accessing content

Abstract: Failed to fetch summary for 2601.18113: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.18113&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2601.20518: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.20518&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2601.21866: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.21866&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.02592: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02592&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as paper content could not be retrieved

Result: Unable to determine results as paper content could not be retrieved

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2602.05474: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.05474&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper fetch

Result: Cannot determine results due to failed paper fetch

Conclusion: Cannot determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2602.11506: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.11506&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.08917: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.08917&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.08206: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08206&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.13165: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13165&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2603.05772: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05772&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2603.05773: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05773&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method cannot be analyzed without access to the paper’s abstract content.

Result: No results can be reported since the paper content is unavailable.

Conclusion: The paper cannot be properly analyzed due to technical limitations preventing access to its content.

Abstract: Failed to fetch summary for 2603.12183: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12183&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Preference-based alignment approach to adapt Japanese SpeechLLMs for speech-worthy outputs (concise, conversational, readily synthesized). Introduces SpokenElyza benchmark derived from ELYZA-tasks-100 with native expert auditory verification

Result: Approach achieves substantial improvement on SpokenElyza benchmark while largely preserving performance on original written-style evaluation

Conclusion: Proposed method successfully adapts Japanese SpeechLLMs for speech-worthy outputs, with SpokenElyza benchmark released to support future Japanese spoken dialog system research

Abstract: SpeechLLMs typically combine ASR-trained encoders with text-based LLM backbones, leading them to inherit written-style output patterns unsuitable for text-to-speech synthesis. This mismatch is particularly pronounced in Japanese, where spoken and written registers differ substantially in politeness markers, sentence-final particles, and syntactic complexity. We propose a preference-based alignment approach to adapt Japanese SpeechLLMs for speech-worthy outputs: text that is concise, conversational, and readily synthesized as natural speech. To rigorously evaluate this task, we introduce SpokenElyza, a benchmark for Japanese speech-worthiness derived from ELYZA-tasks-100 with auditory verification by native experts. Experiments show that our approach achieves substantial improvement on SpokenElyza while largely preserving performance on the original written-style evaluation. We will release SpokenElyza to support future research on Japanese spoken dialog systems.

Method: Two-stage framework: 1) Discriminative masking for coarse separation, 2) Flow matching to refine the output toward target speech. Unlike generative approaches starting from Gaussian noise, this method starts from the masked spectrogram, enabling single-step inference.

Result: Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.

Conclusion: The proposed framework successfully combines discriminative and generative approaches for TSE, achieving high-quality speech extraction with efficient single-step inference.

Abstract: Target speaker extraction (TSE) extracts the target speaker’s voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.

Method: Proposes a dual-stream attacker with parallel encoders for spectral and self-supervised learning features. Uses three-stage training: Stage I establishes foundational speaker-discriminative representations; Stage II leverages shared identity-transformation characteristics of voice conversion and anonymization to build cross-system robustness; Stage III provides lightweight adaptation to target anonymized data.

Result: Results on VoicePrivacy Attacker Challenge dataset show Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10% of target anonymization dataset surpasses current state-of-the-art attackers in terms of Equal Error Rate (EER).

Conclusion: The proposed three-stage training strategy effectively improves attacker generalization across different anonymization systems, with Stage II playing a crucial role in building cross-system robustness through exposure to diverse converted speech.

Abstract: Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.

Method: Uses source-separated stems, physics-informed harmonic modeling, and beat-level audio descriptors to examine melodic, harmonic, and rhythmic relationships across eight recordings with four singers. Implements a commonality-diversity framework combining multi-scale correlation analysis with residual-based detection of structural deviations.

Result: Expressive coordination is predominantly piece-specific rather than corpus-wide. Diversity events systematically align with formal boundaries and textural shifts, demonstrating the approach can identify musically salient reorganizations with minimal human annotation.

Conclusion: The framework offers a generalizable computational strategy for repertoires without notated blueprints, extending Music Performance Analysis into oral-tradition and improvisation-inflected practices.

Abstract: Computational musicology enables systematic analysis of performative and structural traits in recorded music, yet existing approaches remain largely tailored to notated, score-based repertoires. This study advances a methodology for analyzing voice-guitar interaction in Carlos Paredes’s vocal collaborations - an oral-tradition context where compositional and performative layers co-emerge. Using source-separated stems, physics-informed harmonic modelling, and beat-level audio descriptors, we examine melodic, harmonic, and rhythmic relationships across eight recordings with four singers. Our commonality-diversity framework, combining multi-scale correlation analysis with residual-based detection of structural deviations, reveals that expressive coordination is predominantly piece-specific rather than corpus-wide. Diversity events systematically align with formal boundaries and textural shifts, demonstrating that the proposed approach can identify musically salient reorganizations with minimal human annotation. The framework further offers a generalizable computational strategy for repertoires without notated blueprints, extending Music Performance Analysis into oral-tradition and improvisation-inflected practices.

Method: Proposes a flow-matching-based latent refinement framework that learns a continuous vector field to transport inference-time latent samples toward posterior-like latent representations through ODE-based integration before waveform generation, without redesigning the acoustic decoder.

Result: Experimental results on Korean and Chinese singing datasets show improved objective metrics and perceptual quality while maintaining practical synthesis efficiency.

Conclusion: Reducing training-inference latent mismatch is a useful direction for improving expressive singing voice synthesis, and the proposed lightweight latent refinement method is effective and compatible with parallel synthesis backbones.

Abstract: Singing voice synthesis (SVS) aims to generate natural and expressive singing waveforms from symbolic musical scores. In cVAE-based SVS, however, a mismatch arises because the decoder is trained with latent representations inferred from target singing signals, while inference relies on latent representations predicted only from conditioning inputs. This discrepancy can weaken fine expressive acoustic details in the synthesized output. To mitigate this issue, we propose FM-Singer, a flow-matching-based latent refinement framework for cVAE-based singing voice synthesis. Rather than redesigning the acoustic decoder, the proposed method learns a continuous vector field that transports inference-time latent samples toward posterior-like latent representations through ODE-based integration before waveform generation. Because the refinement is performed in latent space, the method remains lightweight and compatible with a strong parallel synthesis backbone. Experimental results on Korean and Chinese singing datasets show that the proposed latent refinement improves objective metrics and perceptual quality while maintaining practical synthesis efficiency. These results suggest that reducing training-inference latent mismatch is a useful direction for improving expressive singing voice synthesis. Code, pre-trained checkpoints, and audio demos are available at https://github.com/alsgur9368/FM-Singer.

Method: The study uses embeddings that capture complementary characteristics: speaker attributes, phonetic content, and semantic meaning. These embeddings are used to select relevant subsets from 100k hours of in-the-wild training data. The approach analyzes how relevance and diversity along these axes affect downstream ASR performance, using CTC-based Conformer models.

Result: Training on a strategically selected 5% subset can exceed the performance of models trained on the full dataset by up to 36.8% relative WER reduction on target domains.

Conclusion: Targeted data selection based on complementary speech characteristics (speaker, phonetic, semantic) is an effective strategy for optimizing specialist ASR models for specific domains, significantly outperforming training on full heterogeneous datasets.

Abstract: Modern ASR systems are typically trained on large-scale pseudo-labeled, in-the-wild data spanning multiple domains. While such heterogeneous data benefit generalist models designed for broad deployment, they pose challenges for specialist models targeting specific domains: specialist models lack the capacity to learn from all available data, and one must pay closer attention to addressing the mismatch between training and test conditions. In this work, we study targeted data selection as a strategy to address these challenges, selecting relevant subsets from 100k hours of in-the-wild training data to optimize performance on target domains. We represent speech samples using embeddings that capture complementary characteristic–speaker attributes, phonetic content, and semantic meaning–and analyze how relevance and diversity along these axes when performing data selection affect downstream ASR performance. Our experiments with CTC-based Conformer models show that training on a strategically selected 5% subset can exceed the performance of models trained on the full dataset by up to 36.8% relative WER reduction on target domains.

Method: Developed C++ Max externals that enable efficient real-time non-linear modal synthesis, with interactive control of physical parameters, custom modal data loading, and multichannel output.

Result: Successfully created open-source software (nlm) that provides real-time non-linear modal synthesis capabilities for strings, membranes, and plates within the Max environment.

Conclusion: nlm makes non-linear modal synthesis more accessible to creative practitioners by integrating advanced physical modeling into a familiar digital audio workstation environment.

Abstract: We present nlm, a set of Max externals that enable efficient real-time non-linear modal synthesis for strings, membranes, and plates. The externals, implemented in C++, offer interactive control of physical parameters, allow the loading of custom modal data, and provide multichannel output. By integrating interactive physical-modelling capabilities into a familiar environment, nlm lowers the barrier for composers, performers, and sound designers to explore the expressive potential of non-linear modal synthesis. The externals are available as open-source software at https://github.com/rodrigodzf/nlm.

Method: Developed yat-product kernel (quadratic alignment + inverse-square proximity), proved its mathematical properties (Mercer kernel, analytic, Lipschitz, self-regularizing), then built Neural Matter Networks using yat-product as the sole non-linearity, replacing traditional blocks with this kernel operation.

Result: NMN-based classifiers match linear baselines on MNIST with bounded prototype evolution and superposition robustness. In language modeling, Aether-GPT2 achieves lower validation loss than GPT-2 with comparable parameters using yat-based attention and MLP blocks.

Conclusion: The framework establishes NMNs as a principled alternative to conventional neural architectures, unifying kernel learning, gradient stability, and information geometry through a single geometrically-grounded kernel operation.

Abstract: We introduce the yat-product, a kernel operator combining quadratic alignment with inverse-square proximity. We prove it is a Mercer kernel, analytic, Lipschitz on bounded domains, and self-regularizing, admitting a unique RKHS embedding. Neural Matter Networks (NMNs) use yat-product as the sole non-linearity, replacing conventional linear-activation-normalization blocks with a single geometrically-grounded operation. This architectural simplification preserves universal approximation while shifting normalization into the kernel itself via the denominator, rather than relying on separate normalization layers. Empirically, NMN-based classifiers match linear baselines on MNIST while exhibiting bounded prototype evolution and superposition robustness. In language modeling, Aether-GPT2 achieves lower validation loss than GPT-2 with a comparable parameter budget while using yat-based attention and MLP blocks. Our framework unifies kernel learning, gradient stability, and information geometry, establishing NMNs as a principled alternative to conventional neural architectures.

Method: Synthesizes principles from Information Theory, Latent Factor Models, and Psychometrics; partitions predictor-space noise into “Predictor Error” and “Structural Uncertainty”; proves asymptotic advantages of high-dimensional error-prone predictors; proposes “Proactive Data-Centric AI” for efficient predictor selection.

Result: Demonstrates that leveraging high-dimensional sets of error-prone predictors asymptotically overcomes both types of noise, while cleaning low-dimensional sets is fundamentally bounded by Structural Uncertainty; shows how “Informative Collinearity” enhances reliability and convergence efficiency.

Conclusion: Redefines data quality from item-level perfection to portfolio-level architecture, providing theoretical rationale for learning from uncurated enterprise “data swamps” and supporting a paradigm shift from “Model Transfer” to “Methodology Transfer” to overcome static generalizability limitations.

Abstract: Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the “Garbage In, Garbage Out” mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space “noise” into “Predictor Error” and “Structural Uncertainty” (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why “Informative Collinearity” (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose “Proactive Data-Centric AI” to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb “rogue” dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI’s focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for “Local Factories” – learning from live, uncurated enterprise “data swamps” – supporting a deployment paradigm shift from “Model Transfer” to “Methodology Transfer’’ to overcome static generalizability limitations.

Method: Proposes MOGP-MMF, a multi-objective genetic programming framework with multi-view multi-level representation (evolutionary, semantic, and structural views), enriched operator set for linear/nonlinear fusion, and improved multi-objective GP with knowledge transfer mechanism.

Result: Outperforms state-of-the-art methods across seven benchmark datasets, particularly in Q8 accuracy and structural integrity. Generates diverse non-dominated solutions for flexible model selection.

Conclusion: MOGP-MMF effectively addresses protein secondary structure prediction through automated feature selection and fusion optimization, providing superior performance and practical flexibility for various application scenarios.

Abstract: Predicting protein secondary structure is essential for understanding protein function and advancing drug discovery. However, the intricate sequence-structure relationship poses significant challenges for accurate modeling. To address these, we propose MOGP-MMF, a multi-objective genetic programming framework that reformulates PSSP as an automated optimization task focused on feature selection and fusion. Specifically, MOGP-MMF introduces a multi-view multi-level representation strategy that integrates evolutionary, semantic, and newly introduced structural views to capture the comprehensive protein folding logic. Leveraging an enriched operator set, the framework evolves both linear and nonlinear fusion functions, effectively capturing high-order feature interactions while reducing fusion complexity. To resolve the accuracy-complexity trade-off, an improved multi-objective GP algorithm is developed, incorporating a knowledge transfer mechanism that utilizes prior evolutionary experience to guide the population toward global optima. Extensive experiments across seven benchmark datasets demonstrate that MOGP-MMF surpasses state-of-the-art methods, particularly in Q8 accuracy and structural integrity. Furthermore, MOGP-MMF generates a diverse set of non-dominated solutions, offering flexible model selection schemes for various practical application scenarios. The source code is available on GitHub: https://github.com/qian-ann/MOGP-MMF/tree/main.

Method: Systematic categorization of generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Benchmarking of existing approaches across four representative BCI paradigms with objective performance comparison.

Result: Provides comprehensive review, methodological taxonomy, benchmark experiments, evaluation metrics, and key applications. Public benchmark codebase available at GitHub repository.

Conclusion: Brain signal generation is crucial for addressing BCI data limitations. Future research should focus on accurate, data-efficient, and privacy-aware BCI systems.

Abstract: Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at https://github.com/wzwvv/DG4BCI.

Method: Feynman agent pipeline: 1) enumerates domain-specific knowledge components (“ideas”), 2) performs code planning, 3) translates ideas into declarative programs, 4) iterates with feedback for visual refinement, 5) renders using Penrose diagramming system with optimization-based layout that preserves visual semantics while injecting randomness.

Result: Generated over 100k well-aligned diagram-caption pairs, created Diagramma benchmark for evaluating visual reasoning in vision-language models. The pipeline enables low-cost, time-efficient diagram authoring with visual consistency and diversity.

Conclusion: Feynman provides scalable solution for generating high-quality vision-language data for multimodal AI systems, addressing data scarcity in visual design applications. The open-source release will benefit research community.

Abstract: Visual design is an essential application of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable diagram generation pipeline built with our agent, Feynman. To create diagrams, Feynman first enumerates domain-specific knowledge components (‘‘ideas’’) and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100k well-aligned diagram-caption pairs. We also curate a visual-language benchmark, Diagramma, from freshly generated data. Diagramma can be used for evaluating the visual reasoning capabilities of vision-language models. We plan to release the dataset, benchmark, and the full agent pipeline as an open-source project.

Method: GER-steer exploits the geometric stability of network representation evolution to refine raw steering vectors. It uses global evolutionary signals to decouple robust semantic intent from orthogonal artifacts without layer-specific tuning.

Result: Extensive evaluations show GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without requiring layer-specific tuning.

Conclusion: GER-steer establishes a universal solution for reliable model alignment through training-free activation engineering that leverages global evolutionary patterns.

Abstract: Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network’s representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.

Method: Geometrically-grounded cognitive manifold with coupled Ricci flow enriched by MDL Drive term derived from first principles, modulated by task-loss gradient, with theoretical guarantees and practical O(N log N) algorithm.

Result: Theoretical proofs of monotonic description length decrease, finite topological phase transitions, universal critical behavior, and empirical validation on synthetic tasks showing robust generalization and autonomous model simplification.

Conclusion: Provides principled path toward more autonomous, generalizable, and interpretable AI by unifying geometric deep learning with information-theoretic principles through MDL-driven optimization.

Abstract: This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \textit{coupled Ricci flow}, enriched with a novel \textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\ref{thm:surgery}, \ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \log N)$ per-iteration complexity (Theorem~\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm’s efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.

Method: The paper establishes computational complexity results, proposes approximating shared common information to construct an approximate POSG model, develops a partially observable multi-agent RL algorithm with quasi-polynomial time/sample complexities, and extends the framework to team-optimal solutions in cooperative POSGs.

Result: The approach enables finding approximate equilibria in quasi-polynomial time under certain assumptions, with both time and sample complexities being quasi-polynomial. The framework also works for team-optimal solutions in cooperative POSGs under structural assumptions.

Conclusion: The study opens possibilities for leveraging different information structures from control theory to develop sample- and computation-efficient partially observable multi-agent RL, demonstrating the importance of information-sharing for tractability.

Abstract: We study provable multi-agent reinforcement learning (RL) in the general framework of partially observable stochastic games (POSGs). To circumvent the known hardness results and the use of computationally intractable oracles, we advocate leveraging the potential \emph{information-sharing} among agents, a common practice in empirical multi-agent RL, and a standard model for multi-agent control systems with communication. We first establish several computational complexity results to justify the necessity of information-sharing, as well as the observability assumption that has enabled quasi-polynomial time and sample single-agent RL with partial observations, for tractably solving POSGs. Inspired by the inefficiency of planning in the ground-truth model, we then propose to further \emph{approximate} the shared common information to construct an approximate model of the POSG, in which an approximate \emph{equilibrium} (of the original POSG) can be found in quasi-polynomial-time, under the aforementioned assumptions. Furthermore, we develop a partially observable multi-agent RL algorithm whose time and sample complexities are \emph{both} quasi-polynomial. Finally, beyond equilibrium learning, we extend our algorithmic framework to finding the \emph{team-optimal solution} in cooperative POSGs, i.e., decentralized partially observable Markov decision processes, a more challenging goal. We establish concrete computational and sample complexities under several structural assumptions of the model. We hope our study could open up the possibilities of leveraging and even designing different \emph{information structures}, a well-studied notion in control theory, for developing both sample- and computation-efficient partially observable multi-agent RL.

Method: Hierarchical decomposition into typed causal primitives across four abstraction layers (physical, functional, event, rule). Uses dual-channel routing network to dynamically compose primitives into differentiable Causal Execution Graphs (CEGs). Implements causal-intervention-driven meta-evolution for autonomous self-improvement.

Result: Significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization across simulated physical and social environments. Provides theoretical guarantees for type-safe composition and universal approximation.

Conclusion: HCP-DCNet offers a principled, scalable, and interpretable architecture for AI systems with human-like causal abstraction and continual self-refinement capabilities, bridging the gap between continuous physical dynamics and discrete symbolic reasoning.

Abstract: The ability to understand and reason about cause and effect – encompassing interventions, counterfactuals, and underlying mechanisms – is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if’’ questions. This paper introduces the \emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.

Method: Proposes a geometric framework interpreting reward parameters as coordinates on a task manifold, showing optimal curricula correspond to geodesics that minimize excess thermodynamic work, with application to temperature annealing in maximum-entropy RL via the “MEW” algorithm.

Result: Develops a principled theoretical framework connecting thermodynamics to RL curriculum learning, providing geometric interpretation of task spaces and deriving optimal curriculum schedules based on minimizing excess work.

Conclusion: The framework successfully connects non-equilibrium thermodynamics to RL curriculum design, offering principled geometric interpretations and practical algorithms for optimal curriculum scheduling.

Abstract: Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, “MEW” (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.

Method: Formulates exploration as maximizing steady-state visitation entropy. Uses spectral characterization via dominant eigenvectors of transition matrix. EVE algorithm avoids rollouts through iterative updates similar to value-based methods. Posterior-policy iteration for unregularized objective.

Result: Proves convergence under standard assumptions. Achieves competitive exploration performance vs rollout-based baselines in deterministic grid-world environments.

Conclusion: EVE provides efficient exploration without explicit rollouts, enabling better data collection for pretraining in reward-free settings.

Abstract: Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.

Method: 1) Train specialist decision trees on functional group features, 2) Verbalize learned predictive rules as natural language for rule-augmented context learning, 3) Introduce rule-consistency test-time scaling that ensembles predictions across diverse rules from Random Forest.

Result: Experiments on 22 ADMET properties from TDC benchmark show TreeKD substantially improves LLM performance, narrowing the gap with state-of-the-art specialist models.

Conclusion: TreeKD advances toward practical generalist models for molecular property prediction by enabling LLMs to leverage structural insights difficult to extract from SMILES strings alone through knowledge distillation from tree-based specialists.

Abstract: Molecular Property Prediction (MPP) is a central task in drug discovery. While Large Language Models (LLMs) show promise as generalist models for MPP, their current performance remains below the threshold for practical adoption. We propose TreeKD, a novel knowledge distillation method that transfers complementary knowledge from tree-based specialist models into LLMs. Our approach trains specialist decision trees on functional group features, then verbalizes their learned predictive rules as natural language to enable rule-augmented context learning. This enables LLMs to leverage structural insights that are difficult to extract from SMILES strings alone. We further introduce rule-consistency, a test-time scaling technique inspired by bagging that ensembles predictions across diverse rules from a Random Forest. Experiments on 22 ADMET properties from the TDC benchmark demonstrate that TreeKD substantially improves LLM performance, narrowing the gap with SOTA specialist models and advancing toward practical generalist models for molecular property prediction.

Method: Introduces Budget-Sensitive Discovery Score (BSDS) with 20 machine-checked theorems, penalizing false discoveries (lambda-weighted FDR) and coverage gaps (gamma-weighted). Budget-averaged form is Discovery Quality Score (DQS). Applied to 39 proposers (ML variants and LLM configurations) on MoleculeNet HIV dataset.

Result: Simple RF-based Greedy-ML outperforms all MLP variants and LLM configurations (DQS = -0.046). No LLM surpasses baseline under zero-shot or few-shot evaluation. Hierarchy generalizes across multiple benchmarks and penalty parameters.

Conclusion: LLMs provide no marginal value over existing trained classifiers for drug discovery candidate selection. BSDS/DQS framework is broadly applicable to budget-constrained selection problems with asymmetric error costs.

Abstract: Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies – a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric – 20 theorems machine-checked by the Lean 4 proof assistant – that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers – 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations – using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.

Method: Extends single-model approach by segmenting training data across increasing numbers of language models, creating ecosystems of models evaluated over ten self-training iterations. Experiments test robustness across different model families, parameter sizes, mixing human- and model-generated data, temperature sampling methods, and scaling effects.

Result: Training a single model on entire dataset improves performance short-term but amplifies collapse long-term. Optimal diversity level increases monotonically with self-training iterations. Effect is robust across experimental settings. Scaling up systems amplifies collapse in homogeneous ecosystems, increasing diversity benefits.

Conclusion: In AI monoculture, need to monitor disagreement among AI systems and incentivize domain- and community-specific models for successful long-term knowledge production. Ecosystem diversity is crucial for mitigating model collapse.

Abstract: As artificial intelligence (AI) becomes more widely used, concerns are growing that model collapse could lead to knowledge collapse, i.e. a degradation to a narrow and inaccurate set of ideas. Prior work has demonstrated single-model collapse, defined as performance decay in an AI model trained on its own outputs. Inspired by ecology, we ask whether increasing AI ecosystem diversity (i.e., the number of distinct models) can mitigate such collapse. To study the effect of diversity on model performance, we extend the single-model approach by segmenting the training data across an increasing number of language models and evaluating the resulting ecosystems of models over ten self-training iterations. We find that training a single model on the entire dataset improves performance only in the short term but amplifies collapse over longer horizons. Specifically, we observe that the optimal diversity level (i.e., the level that maximizes performance) increases monotonically with the number of self-training iterations. The observed effect is robust across various experimental settings, including different model families, parameter sizes, mixing human- and model-generated data, and temperature sampling methods, demonstrating the significance of ecosystem diversity for mitigating collapse. Moreover, our experiments with increased model and dataset sizes indicate that scaling up the system can amplify collapse in highly homogeneous ecosystems, thereby increasing the diversity benefits. In the presence of AI monoculture, our results suggest the need to monitor (dis)agreement among AI systems and to incentivize more domain- and community-specific models to ensure successful knowledge production in the long run.

Method: Proposes NeST-S6, a convolutional selective state-space model with spatial PDE-aware core in a nested learning paradigm. Uses convolutional local spatial mixing feeding a spatial PDE-aware SSM core, with nested-learning long-term memory updated by a learned optimizer when prediction errors indicate unmodeled dynamics.

Result: On Milan mobile-traffic grid dataset at three resolutions, achieves lower errors than Mamba-family baseline in single-step and 6-step autoregressive rollouts. Under drift stress tests, nested memory lowers MAE by 48-65%. Speeds full-grid reconstruction by 32x, reduces MACs by 4.3x, and achieves 61% lower per-pixel RMSE compared to per-pixel scanning models.

Conclusion: NeST-S6 provides an efficient, accurate solution for spatiotemporal grid forecasting in cellular networks, balancing computational efficiency with modeling capability through its nested learning architecture and PDE-aware core.

Abstract: Traffic forecasting in cellular networks is a challenging spatiotemporal prediction problem due to strong temporal dependencies, spatial heterogeneity across cells, and the need for scalability to large network deployments. Traditional cell-specific models incur prohibitive training and maintenance costs, while global models often fail to capture heterogeneous spatial dynamics. Recent spatiotemporal architectures based on attention or graph neural networks improve accuracy but introduce high computational overhead, limiting their applicability in large-scale or real-time settings. We study spatiotemporal grid forecasting, where each time step is a 2D lattice of traffic values, and predict the next grid patch using previous patches. We propose NeST-S6, a convolutional selective state-space model (SSM) with a spatial PDE-aware core, implemented in a nested learning paradigm: convolutional local spatial mixing feeds a spatial PDE-aware SSM core, while a nested-learning long-term memory is updated by a learned optimizer when one-step prediction errors indicate unmodeled dynamics. On the mobile-traffic grid (Milan dataset) at three resolutions (202, 502, 1002), NeST-S6 attains lower errors than a strong Mamba-family baseline in both single-step and 6-step autoregressive rollouts. Under drift stress tests, our model’s nested memory lowers MAE by 48-65% over a no-memory ablation. NeST-S6 also speeds full-grid reconstruction by 32 times and reduces MACs by 4.3 times compared to competitive per-pixel scanning models, while achieving 61% lower per-pixel RMSE.

Method: Theoretical analysis connecting drifting dynamics to Sinkhorn divergence gradient flows, showing how Sinkhorn divergence yields a cross-minus-self structure similar to drifting but with two-sided Sinkhorn scaling. Experimental validation on FFHQ-ALAE and MNIST datasets.

Result: Sinkhorn drifting reduces sensitivity to kernel temperature and improves one-step generative quality. On FFHQ-ALAE, it reduces mean FID from 187.7 to 37.1 and mean latent EMD from 453.3 to 144.4 at lowest temperature settings, while preserving full class coverage on MNIST.

Conclusion: Sinkhorn drifting provides a theoretically grounded connection between drifting dynamics and Sinkhorn divergence gradient flows, resolving identifiability issues and offering practical improvements in generative modeling stability and quality.

Abstract: We establish a theoretical link between the recently proposed “drifting” generative dynamics and gradient flows induced by the Sinkhorn divergence. In a particle discretization, the drift field admits a cross-minus-self decomposition: an attractive term toward the target distribution and a repulsive/self-correction term toward the current model, both expressed via one-sided normalized Gibbs kernels. We show that Sinkhorn divergence yields an analogous cross-minus-self structure, but with each term defined by entropic optimal-transport couplings obtained through two-sided Sinkhorn scaling (i.e., enforcing both marginals). This provides a precise sense in which drifting acts as a surrogate for a Sinkhorn-divergence gradient flow, interpolating between one-sided normalization and full two-sided Sinkhorn scaling. Crucially, this connection resolves an identifiability gap in prior drifting formulations: leveraging the definiteness of the Sinkhorn divergence, we show that zero drift (equilibrium of the dynamics) implies that the model and target measures match. Experiments show that Sinkhorn drifting reduces sensitivity to kernel temperature and improves one-step generative quality, trading off additional training time for a more stable optimization, without altering the inference procedure used by drift methods. These theoretical gains translate to strong low-temperature improvements in practice: on FFHQ-ALAE at the lowest temperature setting we evaluate, Sinkhorn drifting reduces mean FID from 187.7 to 37.1 and mean latent EMD from 453.3 to 144.4, while on MNIST it preserves full class coverage across the temperature sweep. Project page: https://mint-vu.github.io/SinkhornDrifting/

Method: Proposes NeuroLoRA with: 1) Lightweight learnable neuromodulation gate that contextually rescales projection space before expert selection, 2) Contrastive Orthogonality Loss to enforce separation between expert subspaces, 3) Mixture-of-Experts architecture built on frozen random projections.

Result: Outperforms FlyLoRA and other baselines on MMLU, GSM8K, and ScienceQA benchmarks across single-task adaptation, multi-task model merging, and sequential continual learning scenarios while maintaining parameter efficiency.

Conclusion: NeuroLoRA successfully introduces context-aware neuromodulation to LoRA frameworks, improving performance across various adaptation scenarios while preserving computational efficiency of frozen projections.

Abstract: Parameter-Efficient Fine-Tuning (PEFT) techniques, particularly Low-Rank Adaptation (LoRA), have become essential for adapting Large Language Models (LLMs) to downstream tasks. While the recent FlyLoRA framework successfully leverages bio-inspired sparse random projections to mitigate parameter interference, it relies on a static, magnitude-based routing mechanism that is agnostic to input context. In this paper, we propose NeuroLoRA, a novel Mixture-of-Experts (MoE) based LoRA framework inspired by biological neuromodulation – the dynamic regulation of neuronal excitability based on context. NeuroLoRA retains the computational efficiency of frozen random projections while introducing a lightweight, learnable neuromodulation gate that contextually rescales the projection space prior to expert selection. We further propose a Contrastive Orthogonality Loss to explicitly enforce separation between expert subspaces, enhancing both task decoupling and continual learning capacity. Extensive experiments on MMLU, GSM8K, and ScienceQA demonstrate that NeuroLoRA consistently outperforms FlyLoRA and other strong baselines across single-task adaptation, multi-task model merging, and sequential continual learning scenarios, while maintaining comparable parameter efficiency.

Method: The paper demonstrates Hidden State Poisoning attacks that drive the spectral radius ρ(Ā) of discretized transition operators toward zero, collapsing memory from millions to dozens of tokens. They prove an Evasion Existence Theorem showing output-only defenses are insufficient, then introduce SpectralGuard - a real-time monitor tracking spectral stability across all model layers with sub-15ms latency.

Result: SpectralGuard achieves F1=0.961 against non-adaptive attackers and retains F1=0.842 under strongest adaptive settings. Causal interventions and cross-architecture transfer to hybrid SSM-Attention systems confirm spectral monitoring provides effective safety.

Conclusion: Spectral monitoring offers a principled, deployable safety layer for recurrent foundation models by detecting memory-collapsing attacks that evade output-level detection, addressing a critical vulnerability in SSM architectures.

Abstract: State Space Models (SSMs) such as Mamba achieve linear-time sequence processing through input-dependent recurrence, but this mechanism introduces a critical safety vulnerability. We show that the spectral radius rho(A-bar) of the discretized transition operator governs effective memory horizon: when an adversary drives rho toward zero through gradient-based Hidden State Poisoning, memory collapses from millions of tokens to mere dozens, silently destroying reasoning capacity without triggering output-level alarms. We prove an Evasion Existence Theorem showing that for any output-only defense, adversarial inputs exist that simultaneously induce spectral collapse and evade detection, then introduce SpectralGuard, a real-time monitor that tracks spectral stability across all model layers. SpectralGuard achieves F1=0.961 against non-adaptive attackers and retains F1=0.842 under the strongest adaptive setting, with sub-15ms per-token latency. Causal interventions and cross-architecture transfer to hybrid SSM-Attention systems confirm that spectral monitoring provides a principled, deployable safety layer for recurrent foundation models.

Method: Introduced semi-synthetic data augmentation method that generates contexts both descriptive of temporal dynamics and verifiably complementary to numerical histories. This enables massive-scale dataset creation, resulting in CAF-7M corpus with rigorously verified test set.

Result: Demonstrated that semi-synthetic pre-training transfers effectively to real-world evaluation and shows clear evidence of context utilization. Results suggest dataset quality, not architectural limitations, has been the primary bottleneck in context-aided forecasting.

Conclusion: The quality of context data is crucial for multimodal forecasting performance. The proposed semi-synthetic data augmentation approach successfully addresses previous limitations and enables effective context utilization in forecasting models.

Abstract: Context-aided forecasting (CAF) holds promise for integrating domain knowledge and forward-looking information, enabling AI systems to surpass traditional statistical methods. However, recent empirical studies reveal a puzzling gap: multimodal models often fail to outperform their unimodal counterparts. We hypothesize that this underperformance stems from poor context quality in existing datasets, as verification is challenging. To address these limitations, we introduce a semi-synthetic data augmentation method that generates contexts both descriptive of temporal dynamics and verifiably complementary to numerical histories. This approach enables massive-scale dataset creation, resulting in CAF-7M, a corpus of 7 million context-augmented time series windows, including a rigorously verified test set. We demonstrate that semi-synthetic pre-training transfers effectively to real-world evaluation, and show clear evidence of context utilization. Our results suggest that dataset quality, rather than architectural limitations, has been the primary bottleneck in context-aided forecasting.

Method: Find explicit real and complex bases for kernels that respect simpler invariance conditions at a reference point x₀, then use steerability equations to extend to arbitrary points x = g·x₀ for different symmetry groups and tensor types.

Result: Provides ready-to-use kernel bases for various symmetry groups and tensor types, bypassing the need for Clebsch-Gordan coefficient computation and working directly with input/output representations.

Conclusion: The method offers a more accessible approach to designing steerable equivariant CNNs with reduced mathematical complexity, potentially accelerating research and applications in equivariant deep learning.

Abstract: We present an alternative way of solving the steerable kernel constraint that appears in the design of steerable equivariant convolutional neural networks. We find explicit real and complex bases which are ready to use, for different symmetry groups and for feature maps of arbitrary tensor type. A major advantage of this method is that it bypasses the need to numerically or analytically compute Clebsch-Gordan coefficients and works directly with the representations of the input and output feature maps. The strategy is to find a basis of kernels that respect a simpler invariance condition at some point $x_0$, and then \textit{steer} it with the defining equation of steerability to move to some arbitrary point $x=g\cdot x_0$. This idea has already been mentioned in the literature before, but not advanced in depth and with some generality. Here we describe how it works with minimal technical tools to make it accessible for a general audience.

Method: MLNNs integrate Kripke semantics into neural architectures, creating a differentiable “Logic Layer” with Necessity Neurons (□) and Learnable Accessibility (Aθ) that enable reasoning about modal concepts. This bridges symbolic logic and neural networks for financial applications.

Result: Four case studies demonstrate MLNN applications: promoting compliance in trading agents, recovering latent trust networks for market surveillance, encouraging robustness under stress scenarios, and distinguishing statistical belief from verified knowledge to mitigate robo-advisory hallucinations.

Conclusion: MLNNs provide a promising framework to bridge the gap between deep learning’s performance and symbolic logic’s interpretability in finance, enabling differentiable reasoning about regulatory constraints, market dynamics, and knowledge verification.

Abstract: The financial industry faces a critical dichotomy in AI adoption: deep learning often delivers strong empirical performance, while symbolic logic offers interpretability and rule adherence expected in regulated settings. We use Modal Logical Neural Networks (MLNNs) as a bridge between these worlds, integrating Kripke semantics into neural architectures to enable differentiable reasoning about necessity, possibility, time, and knowledge. We illustrate MLNNs as a differentiable ``Logic Layer’’ for finance by mapping core components, Necessity Neurons ($\Box$) and Learnable Accessibility ($A_θ$), to regulatory guardrails, market stress testing, and collusion detection. Four case studies show how MLNN-style constraints can promote compliance in trading agents, help recover latent trust networks for market surveillance, encourage robustness under stress scenarios, and distinguish statistical belief from verified knowledge to help mitigate robo-advisory hallucinations.

Method: Use controlled vision task where Mamba reconstructs images from sequences of image patches; analyze reconstructions at different sequence processing stages; develop length-adaptive Mamba variant to improve performance across training sequence lengths.

Result: Mamba qualitatively adapts its behavior to training sequence length distribution, developing strategies that fail to generalize beyond this range; length-adaptive variant shows improved performance across training sequence lengths.

Conclusion: Provides intuitive perspective on length generalization in Mamba, revealing adaptation strategies that limit generalization, and suggests architectural improvements through length-adaptive variants.

Abstract: Mamba has attracted widespread interest as a general-purpose sequence model due to its low computational complexity and competitive performance relative to transformers. However, its performance can degrade when inference sequence lengths exceed those seen during training. We study this phenomenon using a controlled vision task in which Mamba reconstructs images from sequences of image patches. By analyzing reconstructions at different stages of sequence processing, we reveal that Mamba qualitatively adapts its behavior to the distribution of sequence lengths encountered during training, resulting in strategies that fail to generalize beyond this range. To support our analysis, we introduce a length-adaptive variant of Mamba that improves performance across training sequence lengths. Our results provide an intuitive perspective on length generalization in Mamba and suggest directions for improving the architecture.

Method: Adaptive Conditional Forest Sampling (ACFS) integrates Generalised Random Forests for decision-conditional distribution approximation, CEM-guided global exploration, rank-weighted focused augmentation, surrogate-to-oracle two-stage reranking, and multi-start gradient-based refinement.

Result: ACFS achieves lowest median oracle spectral risk on second benchmark in every configuration (6.0% to 20.0% improvement over GP-BO), reduces cross-replication dispersion by 1.7-2.0 times, and outperforms CEM-SO, SGD-CVaR, and KDE-SO in nearly all settings.

Conclusion: ACFS provides materially improved run-to-run reliability and performance for spectral risk minimization with decision-dependent uncertainty, with ablation and sensitivity analyses supporting the proposed design’s contribution and robustness.

Abstract: Minimising a spectral risk objective, defined as a convex combination of expected cost and Conditional Value-at-Risk (CVaR), is challenging when the uncertainty distribution is decision-dependent, making both surrogate modelling and simulation-based ranking sensitive to tail estimation error. We propose Adaptive Conditional Forest Sampling (ACFS), a four-phase simulation-optimisation framework that integrates Generalised Random Forests for decision-conditional distribution approximation, CEM-guided global exploration, rank-weighted focused augmentation, and surrogate-to-oracle two-stage reranking before multi-start gradient-based refinement. We evaluate ACFS on two structurally distinct data-generating processes: a decision-dependent Student-t copula and a Gaussian copula with log-normal marginals, across three penalty-weight configurations and 100 replications per setting. ACFS achieves the lowest median oracle spectral risk on the second benchmark in every configuration, with median gaps over GP-BO ranging from 6.0% to 20.0%. On the first benchmark, ACFS and GP-BO are statistically indistinguishable in median objective, but ACFS reduces cross-replication dispersion by approximately 1.8 to 1.9 times on the first benchmark and 1.7 to 2.0 times on the second, indicating materially improved run-to-run reliability. ACFS also outperforms CEM-SO, SGD-CVaR, and KDE-SO in nearly all settings, while ablation and sensitivity analyses support the contribution and robustness of the proposed design.

Method: Byz-NSGDM combines momentum normalization with Byzantine-robust aggregation enhanced by Nearest Neighbor Mixing (NNM). The algorithm normalizes stochastic gradients with momentum to handle (L₀,L₁)-smoothness while using NNM-enhanced aggregation to filter out Byzantine workers’ contributions.

Result: Theoretical analysis proves O(K⁻¹/⁴) convergence rate up to a Byzantine bias floor proportional to robustness coefficient and gradient heterogeneity. Experiments on heterogeneous MNIST classification, synthetic (L₀,L₁)-smooth optimization, and character-level language modeling with GPT show effectiveness against various Byzantine attacks.

Conclusion: Byz-NSGDM provides a robust distributed optimization method that handles both (L₀,L₁)-smoothness and Byzantine adversaries, with theoretical guarantees and empirical validation across diverse tasks and attack strategies.

Abstract: We consider distributed optimization under Byzantine attacks in the presence of $(L_0,L_1)$-smoothness, a generalization of standard $L$-smoothness that captures functions with state-dependent gradient Lipschitz constants. We propose Byz-NSGDM, a normalized stochastic gradient descent method with momentum that achieves robustness against Byzantine workers while maintaining convergence guarantees. Our algorithm combines momentum normalization with Byzantine-robust aggregation enhanced by Nearest Neighbor Mixing (NNM) to handle both the challenges posed by $(L_0,L_1)$-smoothness and Byzantine adversaries. We prove that Byz-NSGDM achieves a convergence rate of $O(K^{-1/4})$ up to a Byzantine bias floor proportional to the robustness coefficient and gradient heterogeneity. Experimental validation on heterogeneous MNIST classification, synthetic $(L_0,L_1)$-smooth optimization, and character-level language modeling with a small GPT model demonstrates the effectiveness of our approach against various Byzantine attack strategies. An ablation study further shows that Byz-NSGDM is robust across a wide range of momentum and learning rate choices.

Method: Multimodal learning framework coupling graph network simulator for Lagrangian tracer-particle motion with 3D U-Net for voxelized interface evolution. Uses imaged pore geometry as boundary constraint, with coupled iterative updates at each time step. Trained autoregressively on experimental sequences in capillary-dominated conditions.

Result: The learned surrogate captures transient flow perturbations and abrupt interface rearrangements (Haines jumps) over seconds of physical time, reducing hour-to-day scale direct numerical simulations to seconds of inference while maintaining experimental accuracy.

Conclusion: The framework enables rapid, experimentally informed “digital experiments” for replicating pore-scale physics, offering efficient exploration of injection conditions and pore-geometry effects relevant to subsurface carbon and hydrogen storage applications.

Abstract: Multiphase flow in porous media underpins subsurface energy and environmental technologies, including geological CO$_2$ storage and underground hydrogen storage, yet pore-scale dynamics in realistic three-dimensional materials remain difficult to characterize and predict. Here we introduce a multimodal learning framework that infers multiphase pore-scale flow directly from time-resolved four-dimensional (4D) micro-velocimetry measurements. The model couples a graph network simulator for Lagrangian tracer-particle motion with a 3D U-Net for voxelized interface evolution. The imaged pore geometry serves as a boundary constraint to the flow velocity and the multiphase interface predictions, which are coupled and updated iteratively at each time step. Trained autoregressively on experimental sequences in capillary-dominated conditions ($Ca\approx10^{-6}$), the learned surrogate captures transient, nonlocal flow perturbations and abrupt interface rearrangements (Haines jumps) over rollouts spanning seconds of physical time, while reducing hour-to-day–scale direct numerical simulations to seconds of inference. By providing rapid, experimentally informed predictions, the framework opens a route to ‘‘digital experiments’’ to replicate pore-scale physics observed in multiphase flow experiments, offering an efficient tool for exploring injection conditions and pore-geometry effects relevant to subsurface carbon and hydrogen storage.

Method: Analyzed per-timestep training losses and identified a U-shaped difficulty profile with persistent errors near boundary regimes. Proposed Curriculum Sampling: a two-phase schedule that begins with middle-biased sampling for rapid structure learning, then switches to Uniform sampling for boundary refinement to resolve fine details.

Result: On CIFAR-10, Curriculum Sampling improved the best FID from 3.85 (Uniform) to 3.22 while reaching peak performance at 100k rather than 150k training steps. This demonstrates both faster convergence and better final quality.

Conclusion: Timestep sampling should be treated as an evolving curriculum rather than a fixed hyperparameter. Curriculum Sampling effectively addresses the speed-quality trade-off in Flow Matching models by adaptively changing sampling strategies during training.

Abstract: Timestep sampling $p(t)$ is a central design choice in Flow Matching models, yet common practice increasingly favors static middle-biased distributions (e.g., Logit-Normal). We show that this choice induces a speed–quality trade-off: middle-biased sampling accelerates early convergence but yields worse asymptotic fidelity than Uniform sampling. By analyzing per-timestep training losses, we identify a U-shaped difficulty profile with persistent errors near the boundary regimes, implying that under-sampling the endpoints leaves fine details unresolved. Guided by this insight, we propose \textbf{Curriculum Sampling}, a two-phase schedule that begins with middle-biased sampling for rapid structure learning and then switches to Uniform sampling for boundary refinement. On CIFAR-10, Curriculum Sampling improves the best FID from $3.85$ (Uniform) to $3.22$ while reaching peak performance at $100$k rather than $150$k training steps. Our results highlight that timestep sampling should be treated as an evolving curriculum rather than a fixed hyperparameter.

Method: Analyzed 5,000-prompt best-of-4 benchmark from Chatbot Arena, comparing global correlation vs within-prompt correlation, tie rates, and recovery metrics. Conducted matched-pair best-of-2 audit with explicit pairwise judging.

Result: Judge with moderate global correlation (r=0.47) captured only 21.0% of potential improvement; within-prompt correlation was much lower (r_within=0.27) with 67% tie rate. Pairwise judging improved recovery from 21.1% to 61.2%.

Conclusion: For judge-based selection tasks, evaluation should focus on within-prompt signal, tie rates, and recovery/top-1 accuracy rather than global agreement alone. Pairwise judging significantly improves selection performance.

Abstract: Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.

Method: TERMINATOR leverages the predictability of first answer arrivals to create a dataset of optimal reasoning lengths. It uses these positions to train an early-exit strategy that determines when to stop reasoning, preventing unnecessary computation.

Result: TERMINATOR achieves significant reductions in Chain-of-Thought lengths: 14%-55% on average across four challenging datasets (MATH-500, AIME 2025, HumanEval, and GPQA) while outperforming current state-of-the-art methods.

Conclusion: The proposed TERMINATOR approach effectively mitigates overthinking in Large Reasoning Models by predicting optimal reasoning lengths, leading to substantial efficiency gains without compromising performance.

Abstract: Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM’s final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.

Method: Construct stationary surrogate action set and use delayed-update scheme to control bias from nonstationary conditional context distributions; provide phased algorithm for unknown transition distributions that learns surrogate mapping online.

Result: Obtain high-probability worst-case regret bound matching underlying linear bandit oracle, with only lower-order dependence on mixing time.

Conclusion: Successfully extend reduction approach to Markovian contextual linear bandits, enabling application of mature linear bandit techniques to temporally correlated settings.

Abstract: Recent work shows that when contexts are drawn i.i.d., linear contextual bandits can be reduced to single-context linear bandits. This ``contexts are cheap” perspective is highly advantageous, as it allows for sharper finite-time analyses and leverages mature techniques from the linear bandit literature, such as those for misspecification and adversarial corruption. Motivated by applications with temporally correlated availability, we extend this perspective to Markovian contextual linear bandits, where the action set evolves via an exogenous Markov chain. Our main contribution is a reduction that applies under uniform geometric ergodicity. We construct a stationary surrogate action set to solve the problem using a standard linear bandit oracle, employing a delayed-update scheme to control the bias induced by the nonstationary conditional context distributions. We further provide a phased algorithm for unknown transition distributions that learns the surrogate mapping online. In both settings, we obtain a high-probability worst-case regret bound matching that of the underlying linear bandit oracle, with only lower-order dependence on the mixing time.

Method: Analyzes EQML feasibility through two implementation pathways: 1) hybrid workflows with remote QPU offloading, and 2) embedded QPU co-processors. Also considers quantum-inspired classical methods on embedded processors/FPGAs as a bridge technology.

Result: Identifies dominant barriers: latency, data encoding overhead, NISQ noise, tooling mismatch, and energy. Maps these to engineering directions in interface design, control electronics, power management, verification, and security.

Conclusion: EQML is technically feasible in limited experimental forms, with quantum-inspired classical methods serving as a practical near-term bridge. Responsible deployment requires adversarial evaluation and governance practices for edge AI systems.

Abstract: Embedded quantum machine learning (EQML) seeks to bring quantum machine learning (QML) capabilities to resource-constrained edge platforms such as IoT nodes, wearables, drones, and cyber-physical controllers. In 2026, EQML is technically feasible only in limited and highly experimental forms: (i) hybrid workflows where an embedded device performs sensing and classical processing while offloading a narrowly scoped quantum subroutine to a remote quantum processing unit (QPU) or nearby quantum appliance, and (ii) early-stage “embedded QPU” concepts in which a compact quantum co-processor is integrated with classical control hardware. A practical bridge is quantum-inspired machine learning and optimisation on classical embedded processors and FPGAs. This paper analyses feasibility from a circuits-and-systems perspective aligned with the academic community, formalises two implementation pathways, identifies the dominant barriers (latency, data encoding overhead, NISQ noise, tooling mismatch, and energy), and maps them to concrete engineering directions in interface design, control electronics, power management, verification, and security. We also argue that responsible deployment requires adversarial evaluation and governance practices that are increasingly necessary for edge AI systems.

Method: Developed low-order linear surrogates (32-dimensional) to approximate transformer depth dynamics, tested across multiple tasks (toxicity, irony, hate speech, sentiment) on GPT-2 models, and analyzed scaling properties across the GPT-2 family.

Result: Found near-perfect agreement between linear surrogates and full GPT-2-large models in layerwise sensitivity profiles, discovered scaling principle where linear surrogate accuracy improves monotonically with model size, and demonstrated efficient multi-layer interventions using linear approximations.

Conclusion: Large language models exhibit emergent low-order linear depth dynamics that become more pronounced with scaling, providing a systems-theoretic foundation for analyzing and controlling transformer models through principled linear approximations.

Abstract: Large language models are often viewed as high-dimensional nonlinear systems and treated as black boxes. Here, we show that transformer depth dynamics admit accurate low-order linear surrogates within context. Across tasks including toxicity, irony, hate speech and sentiment, a 32-dimensional linear surrogate reproduces the layerwise sensitivity profile of GPT-2-large with near-perfect agreement, capturing how the final output shifts under additive injections at each layer. We then uncover a surprising scaling principle: for a fixed-order linear surrogate, agreement with the full model improves monotonically with model size across the GPT-2 family. This linear surrogate also enables principled multi-layer interventions that require less energy than standard heuristic schedules when applied to the full model. Together, our results reveal that as language models scale, low-order linear depth dynamics emerge within contexts, offering a systems-theoretic foundation for analyzing and controlling them.

Method: Introduces CALF (Communication-Aware Learning Framework) which trains policies under realistic network models during simulation, explicitly modeling communication constraints

Result: Network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines, validated through distributed policy deployments across heterogeneous hardware

Conclusion: Network conditions are a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomization

Abstract: Distributed reinforcement learning policies face network delays, jitter, and packet loss when deployed across edge devices and cloud servers. Standard RL training assumes zero-latency interaction, causing severe performance degradation under realistic network conditions. We introduce CALF (Communication-Aware Learning Framework), which trains policies under realistic network models during simulation. Systematic experiments demonstrate that network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines. Distributed policy deployments across heterogeneous hardware validate that explicitly modelling communication constraints during training enables robust real-world execution. These findings establish network conditions as a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomisation.

Method: Formulates diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory, derives an exact unbiased policy gradient that decomposes over denoising steps, uses entropy-guided step selection for policy updates, and estimates intermediate advantages using one-step denoising rewards.

Result: Achieves state-of-the-art results on coding and logical reasoning benchmarks, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for diffusion language models.

Conclusion: The proposed method provides an effective RL framework for diffusion language models that avoids biased approximations and costly rollouts while maintaining computational efficiency.

Abstract: Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at https://github.com/vishnutez/egspo-dllm-rl.

Method: DDMM uses a learning algorithm that assigns weights to each anchor-positive sample pair based on Euclidean distance, then learns the differences within pairs weighted by these weights, enabling learning of minute differences while sampling from entire time series.

Result: DDMM significantly outperformed state-of-the-art time series representation learning methods on the Pulp-and-paper mill dataset and showed further accuracy improvements when combined with existing feature extraction methods.

Conclusion: DDMM effectively improves retrieval accuracy for industrial time series data by learning minute state differences, demonstrating practical value for industrial plant applications.

Abstract: We propose the Deep Distance Measurement Method (DDMM) to improve retrieval accuracy in unsupervised multivariate time series similarity retrieval. DDMM enables learning of minute differences within states in the entire time series and thereby recognition of minute differences between states, which are of interest to users in industrial plants. To achieve this, DDMM uses a learning algorithm that assigns a weight to each pair of an anchor and a positive sample, arbitrarily sampled from the entire time series, based on the Euclidean distance within the pair and learns the differences within the pairs weighted by the weights. This algorithm allows both learning minute differences within states and sampling pairs from the entire time series. Our empirical studies showed that DDMM significantly outperformed state-of-the-art time series representation learning methods on the Pulp-and-paper mill dataset and demonstrated the effectiveness of DDMM in industrial plants. Furthermore, we showed that accuracy can be further improved by linking DDMM with existing feature extraction methods through experiments with the combined model.

Method: Curvature-Aware Heterogeneous Federated Pruning (CA-HFP) framework where each client performs device-specific structured pruning guided by curvature-informed significance scores, then maps compact submodels back to common global parameter space via lightweight reconstruction.

Result: Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet show CA-HFP preserves accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning baselines.

Conclusion: CA-HFP provides a practical solution for efficient federated learning on heterogeneous edge devices through personalized pruning with theoretical convergence guarantees and empirical effectiveness.

Abstract: Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.

Method: Model embedding evolution under Langevin dynamics on a compact Riemannian manifold, analyze under mild smoothness and energy-barrier assumptions, and apply classical simulated annealing theory to contrastive learning.

Result: Slow logarithmic inverse-temperature schedules ensure convergence in probability to globally optimal representations, while faster schedules risk trapping in suboptimal minima, establishing a link between contrastive learning and simulated annealing.

Conclusion: The analysis provides a principled theoretical basis for understanding and tuning temperature schedules in contrastive learning, connecting it to well-established simulated annealing theory.

Abstract: The InfoNCE loss in contrastive learning depends critically on a temperature parameter, yet its dynamics under fixed versus annealed schedules remain poorly understood. We provide a theoretical analysis by modeling embedding evolution under Langevin dynamics on a compact Riemannian manifold. Under mild smoothness and energy-barrier assumptions, we show that classical simulated annealing guarantees extend to this setting: slow logarithmic inverse-temperature schedules ensure convergence in probability to a set of globally optimal representations, while faster schedules risk becoming trapped in suboptimal minima. Our results establish a link between contrastive learning and simulated annealing, providing a principled basis for understanding and tuning temperature schedules.

Method: Empirical scaling law analysis on canonical nonlinear PDEs using single-layer PINNs, investigating the relationship between network width, nonlinearity, and solution error through systematic experiments.

Result: Identified dual optimization failure: baseline pathology where error doesn’t decrease with width even at fixed nonlinearity, and compounding pathology where nonlinearity worsens this. Found scaling follows complex non-separable relationships rather than simple power laws.

Conclusion: Optimization, not approximation capacity, is the primary bottleneck in PINNs scaling. The failure is consistent with spectral bias where networks struggle with high-frequency components intensified by nonlinearity. Proposed methodology to measure these complex scaling effects.

Abstract: We establish empirical scaling laws for Single-Layer Physics-Informed Neural Networks on canonical nonlinear PDEs. We identify a dual optimization failure: (i) a baseline pathology, where the solution error fails to decrease with network width, even at fixed nonlinearity, falling short of theoretical approximation bounds, and (ii) a compounding pathology, where this failure is exacerbated by nonlinearity. We provide quantitative evidence that a simple separable power law is insufficient, and that the scaling behavior is governed by a more complex, non-separable relationship. This failure is consistent with the concept of spectral bias, where networks struggle to learn the high-frequency solution components that intensify with nonlinearity. We show that optimization, not approximation capacity, is the primary bottleneck, and propose a methodology to empirically measure these complex scaling effects.

Method: Proposes Swap-guided Preference Learning (SPL) with three components: 1) swap-guided base regularization using fictitious swap annotators, 2) Preferential Inverse Autoregressive Flow (P-IAF) for richer latent representations, and 3) adaptive latent conditioning.

Result: SPL mitigates posterior collapse, enriches user-specific latent variables, and improves preference prediction compared to baseline methods.

Conclusion: SPL successfully addresses posterior collapse in preference learning, enabling better personalization in RLHF systems by preserving user-specific preferences through novel swap-guided regularization techniques.

Abstract: Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at https://github.com/cobang0111/SPL

Method: Bridges GNNs with control theory using integer- and fractional-order Lyapunov stability. Proposes an adaptive, learnable Lyapunov function with a novel projection mechanism that maps the network’s state into a stable space, constraining the underlying feature-update dynamics.

Result: Extensive experiments show that Lyapunov-stable graph neural flows substantially outperform base neural flows and state-of-the-art baselines across standard benchmarks and various adversarial attack scenarios.

Conclusion: The framework provides theoretically provable stability guarantees and is orthogonal to existing defenses, allowing integration with techniques like adversarial training for cumulative robustness.

Abstract: Graph Neural Networks (GNNs) are highly vulnerable to adversarial perturbations in both topology and features, making the learning of robust representations a critical challenge. In this work, we bridge GNNs with control theory to introduce a novel defense framework grounded in integer- and fractional-order Lyapunov stability. Unlike conventional strategies that rely on resource-heavy adversarial training or data purification, our approach fundamentally constrains the underlying feature-update dynamics of the GNN. We propose an adaptive, learnable Lyapunov function paired with a novel projection mechanism that maps the network’s state into a stable space, thereby offering theoretically provable stability guarantees. Notably, this mechanism is orthogonal to existing defenses, allowing for seamless integration with techniques like adversarial training to achieve cumulative robustness. Extensive experiments demonstrate that our Lyapunov-stable graph neural flows substantially outperform base neural flows and state-of-the-art baselines across standard benchmarks and various adversarial attack scenarios.

Method: Proposes Consensus Aggregation for Policy Optimization (CAPO) which runs K PPO replicas on the same batch with different minibatch shuffling orders, then aggregates them. Studies two aggregation spaces: Euclidean parameter space and natural parameter space via logarithmic opinion pool.

Result: CAPO outperforms PPO and compute-matched deeper baselines by up to 8.6x on continuous control tasks under fixed sample budgets. Natural parameter space aggregation provably achieves higher KL-penalized surrogate and tighter trust region compliance than mean expert.

Conclusion: Policy optimization can be improved by optimizing wider rather than deeper without additional environment interactions. Consensus aggregation effectively redirects compute from depth to width, addressing the optimization-depth dilemma in PPO.

Abstract: Proximal policy optimization (PPO) approximates the trust region update using multiple epochs of clipped SGD. Each epoch may drift further from the natural gradient direction, creating path-dependent noise. To understand this drift, we can use Fisher information geometry to decompose policy updates into signal (the natural gradient projection) and waste (the Fisher-orthogonal residual that consumes trust region budget without first-order surrogate improvement). Empirically, signal saturates but waste grows with additional epochs, creating an optimization-depth dilemma. We propose Consensus Aggregation for Policy Optimization (CAPO), which redirects compute from depth to width: $K$ PPO replicates are optimized on the same batch, differing only in minibatch shuffling order, and then aggregated into a consensus. We study aggregation in two spaces: Euclidean parameter space, and the natural parameter space of the policy distribution via the logarithmic opinion pool. In natural parameter space, the consensus provably achieves higher KL-penalized surrogate and tighter trust region compliance than the mean expert; parameter averaging inherits these guarantees approximately. On continuous control tasks, CAPO outperforms PPO and compute-matched deeper baselines under fixed sample budgets by up to 8.6x. CAPO demonstrates that policy optimization can be improved by optimizing wider, rather than deeper, without additional environment interactions.

Method: Analyzes distributional Bellman dynamics directly at the level of cumulative distribution functions (CDFs), treating Cramér geometry as intrinsic analytical setting. Shows Bellman update acts affinely on CDFs and linearly on differences between CDFs. Constructs family of regularized spectral Hilbert representations that realize CDF-level geometry by exact conjugation without modifying underlying Bellman dynamics.

Result: Provides intrinsic formulation clarifying operator structure underlying distributional Bellman updates, with contraction property yielding uniform bound on linear action. Regularization affects only geometry and vanishes in zero-regularization limit, recovering native Cramér metric.

Conclusion: The framework clarifies operator structure underlying distributional Bellman updates and provides foundation for further functional and operator-theoretic analyses in distributional reinforcement learning.

Abstract: Distributional reinforcement learning (DRL) studies the evolution of full return distributions under Bellman updates rather than focusing on expected values. A classical result is that the distributional Bellman operator is contractive under the Cramér metric, which corresponds to an $L^2$ geometry on differences of cumulative distribution functions (CDFs). While this contraction ensures stability of policy evaluation, existing analyses remain largely metric, focusing on contraction properties without elucidating the structural action of the Bellman update on distributions. In this work, we analyse distributional Bellman dynamics directly at the level of CDFs, treating the Cramér geometry as the intrinsic analytical setting. At this level, the Bellman update acts affinely on CDFs and linearly on differences between CDFs, and its contraction property yields a uniform bound on this linear action. Building on this intrinsic formulation, we construct a family of regularised spectral Hilbert representations that realise the CDF-level geometry by exact conjugation, without modifying the underlying Bellman dynamics. The regularisation affects only the geometry and vanishes in the zero-regularisation limit, recovering the native Cramér metric. This framework clarifies the operator structure underlying distributional Bellman updates and provides a foundation for further functional and operator-theoretic analyses in DRL.

Method: Proposes IE-CL framework with two components: learnable transformation for entropy generation and encoder regularizer for entropy preservation, optimizing entropy gain between augmented views.

Result: Experiments on CIFAR-10/100, STL-10, and ImageNet show consistent performance improvements under small-batch settings, with modules being easily integrated into existing frameworks.

Conclusion: IE-CL offers a new theoretical perspective on contrastive learning that successfully bridges information theory principles with practical performance gains.

Abstract: Contrastive learning has achieved remarkable success in self-supervised representation learning, often guided by information-theoretic objectives such as mutual information maximization. Motivated by the limitations of static augmentations and rigid invariance constraints, we propose IE-CL (Incremental-Entropy Contrastive Learning), a framework that explicitly optimizes the entropy gain between augmented views while preserving semantic consistency. Our theoretical framework reframes the challenge by identifying the encoder as an information bottleneck and proposes a joint optimization of two components: a learnable transformation for entropy generation and an encoder regularizer for its preservation. Experiments on CIFAR-10/100, STL-10, and ImageNet demonstrate that IE-CL consistently improves performance under small-batch settings. Moreover, our core modules can be seamlessly integrated into existing frameworks. This work bridges theoretical principles and practice, offering a new perspective in contrastive learning.

Method: FastDSAC introduces Dimension-wise Entropy Modulation (DEM) to dynamically redistribute exploration budget and enforce diversity, plus a continuous distributional critic to ensure value fidelity and mitigate high-dimensional value overestimation.

Result: Extensive evaluations on HumanoidBench show stochastic policies can match or outperform deterministic baselines, achieving 180% and 400% gains on challenging Basketball and Balance Hard tasks.

Conclusion: Rigorously designed stochastic policies with proper entropy modulation and value estimation can effectively handle high-dimensional continuous control, challenging the prevailing preference for deterministic approaches.

Abstract: Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a formidable challenge, as the ``curse of dimensionality’’ induces severe exploration inefficiency and training instability in expansive action spaces. Consequently, recent high-throughput paradigms have largely converged on deterministic policy gradients combined with massive parallel simulation. We challenge this compromise with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget and enforce diversity, alongside a continuous distributional critic tailored to ensure value fidelity and mitigate high-dimensional value overestimation. Extensive evaluations on HumanoidBench and other continuous control tasks demonstrate that rigorously designed stochastic policies can consistently match or outperform deterministic baselines, achieving notable gains of 180% and 400% on the challenging \textit{Basketball} and \textit{Balance Hard} tasks.

Method: Proposes OnlineSpec framework that treats speculative decoding as an online learning problem. Uses dynamic regret minimization to link online learning performance to acceleration rate. Implements two algorithms: 1) optimistic online learning that reuses historical gradients as predictive hints, and 2) online ensemble learning that dynamically maintains multiple draft models.

Result: Achieves up to 24% speedup over seven benchmarks and three foundation models, with theoretical justifications for improved acceleration rates.

Conclusion: OnlineSpec successfully leverages the natural feedback loop in speculative decoding to continuously adapt draft models, significantly improving inference acceleration through online learning techniques.

Abstract: Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative “draft commits-feedback provides-draft adapts” evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system’s acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.

Method: Proxy-modelled Bayesian optimization (px-BO) with human-AI teaming: 1) Human agents vote on preferences between experimental outcomes, 2) Bradley-Terry model transforms votes into proxy objective function, 3) Proxy model acts as AI agent for surrogate votes, 4) Periodic human validation and correction of proxy model.

Result: The px-BO framework demonstrated improved search performance on simulated and BEPS data from PTO samples compared to traditional data-driven exploration, providing better domain expert control.

Conclusion: Human-AI teaming via px-BO enables accelerated and meaningful material space exploration by combining human expertise with AI efficiency, addressing limitations of purely data-driven approaches.

Abstract: Optimization for different tasks like material characterization, synthesis, and functional properties for desired applications over multi-dimensional control parameters need a rapid strategic search through active learning such as Bayesian optimization (BO). However, such high-dimensional experimental physical descriptors are complex and noisy, from which realization of a low-dimensional mathematical scalar metrics or objective functions can be erroneous. Moreover, in traditional purely data-driven autonomous exploration, such objective functions often ignore the subtle variation and key features of the physical descriptors, thereby can fail to discover unknown phenomenon of the material systems. To address this, here we present a proxy-modelled Bayesian optimization (px-BO) via on-the-fly teaming between human and AI agents. Over the loop of BO, instead of defining a mathematical objective function directly from the experimental data, we introduce a voting system on the fly where the new experimental outcome will be compared with existing experiments, and the human agents will choose the preferred samples. These human-guided comparisons are then transformed into a proxy-based objective function via fitting Bradley-Terry (BT) model. Then, to minimize human interaction, this iteratively trained proxy model also acts as an AI agent for future surrogate human votes. Finally, these surrogate votes are periodically validated by human agents, and the corrections are then learned by the proxy model on-the-fly. We demonstrated the performance of the proposed px-BO framework into simulated and BEPS data generated from PTO sample. We find that our approach provided better control of the domain experts for an improved search over traditional data-driven exploration, thus, signifies the importance of human-AI teaming in an accelerated and meaningful material space exploration.

Method: BAVT models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. It uses a budget-conditioned node selection mechanism that scales node values by the remaining resource ratio, transitioning from exploration to exploitation as budget depletes. A residual value predictor scores relative progress rather than absolute state quality to combat LLM overconfidence.

Result: BAVT consistently outperforms parallel sampling baselines on four multi-hop QA benchmarks across two model families. Under strict low-budget constraints, BAVT surpasses baseline performance at 4× the resource allocation, showing intelligent budget management outperforms brute-force compute scaling.

Conclusion: BAVT provides a principled, training-free framework for budget-aware LLM agents with theoretical convergence guarantees, demonstrating that intelligent resource management is more effective than simply scaling compute.

Abstract: Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-ε$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.

Method: Uses deep learning diffusion model with multi-step autoregressive objective for long-rollout stability, multi-scale graph transformer architecture for complex geometries, and incorporates uncertainty estimation for adaptive sensor placement and diffusion posterior sampling for data assimilation.

Result: Demonstrated on 2D homogeneous/isotropic turbulence and backward-facing step flow, showing utility in forecasting, adaptive sensor placement, and data assimilation for high-dimensional chaotic systems.

Conclusion: The framework provides a unified platform for probabilistic forecasting, sensor placement, and data assimilation in chaotic systems, addressing limitations of deterministic models.

Abstract: High-fidelity numerical simulations of chaotic, high dimensional nonlinear dynamical systems are computationally expensive, necessitating the development of efficient surrogate models. Most surrogate models for such systems are deterministic, for example when neural operators are involved. However, deterministic models often fail to capture the intrinsic distributional uncertainty of chaotic systems. This work presents a surrogate modeling formulation that leverages generative machine learning, where a deep learning diffusion model is used to probabilistically forecast turbulent flows over long horizons. We introduce a multi-step autoregressive diffusion objective that significantly enhances long-rollout stability compared to standard single-step training. To handle complex, unstructured geometries, we utilize a multi-scale graph transformer architecture incorporating diffusion preconditioning and voxel-grid pooling. More importantly, our modeling framework provides a unified platform that also predicts spatiotemporally important locations for sensor placement, either via uncertainty estimates or through an error-estimation module. Finally, the observations of the ground truth state at these dynamically varying sensor locations are assimilated using diffusion posterior sampling requiring no retraining of the surrogate model. We present our methodology on two-dimensional homogeneous and isotropic turbulence and for a flow over a backwards-facing step, demonstrating its utility in forecasting, adaptive sensor placement, and data assimilation for high dimensional chaotic systems.

Method: Proposes expert replacing paradigm that replaces redundant experts with parameter-efficient modules. LightMoE enhances this with: 1) adaptive expert selection, 2) hierarchical expert construction, and 3) annealed recovery strategy for low training costs.

Result: Matches LoRA fine-tuning performance at 30% compression ratio. At 50% compression rate, outperforms existing methods with average 5.6% performance improvement across five diverse tasks.

Conclusion: LightMoE achieves superior balance among memory efficiency, training efficiency, and model performance for MoE-based LLMs through expert replacing paradigm.

Abstract: Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.

Method: Proposes RetroReasoner with two-stage training: 1) Supervised fine-tuning using SyntheticRetro framework that generates structured disconnection rationales alongside reactant predictions, 2) Reinforcement learning using round-trip accuracy as reward where predicted reactants are passed through a forward synthesis model and rewarded when the forward-predicted product matches the original input product.

Result: Experimental results show RetroReasoner outperforms prior baselines and generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.

Conclusion: RetroReasoner successfully incorporates chemists’ strategic reasoning into retrosynthesis prediction through structured rationale generation and reinforcement learning, improving performance especially on difficult cases.

Abstract: Retrosynthesis prediction is a core task in organic synthesis that aims to predict reactants for a given product molecule. Traditionally, chemists select a plausible bond disconnection and derive corresponding reactants, which is time-consuming and requires substantial expertise. While recent advancements in molecular large language models (LLMs) have made progress, many methods either predict reactants without strategic reasoning or conduct only a generic product analysis, rather than reason explicitly about bond-disconnection strategies that logically lead to the choice of specific reactants. To overcome these limitations, we propose RetroReasoner, a retrosynthetic reasoning model that leverages chemists’ strategic thinking. RetroReasoner is trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we introduce SyntheticRetro, a framework that generates structured disconnection rationales alongside reactant predictions. In the case of RL, we apply a round-trip accuracy as reward, where predicted reactants are passed through a forward synthesis model, and predictions are rewarded when the forward-predicted product matches the original input product. Experimental results show that RetroReasoner not only outperforms prior baselines but also generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.

Method: Introduces Sobolev-Ricci Curvature (SRC) induced by Sobolev transport geometry, which admits efficient evaluation via a tree-metric Sobolev structure on neighborhood measures. Demonstrates consistency with classical transport curvature and applies SRC in two pipelines: Sobolev-Ricci Flow for reweighting and curvature-guided edge pruning for manifold structure preservation.

Result: SRC recovers Ollivier-Ricci curvature on trees with length measure, vanishes in the Dirac limit matching flat case of measure-theoretic Ricci curvature, and provides a transport-based foundation for scalable curvature-driven graph transformation and manifold-oriented pruning.

Conclusion: SRC offers an efficient, theoretically grounded curvature primitive for graph analysis and transformation, with applications in network geometry manipulation and manifold structure preservation through curvature-guided operations.

Abstract: Ricci curvature is a fundamental concept in differential geometry for encoding local geometric structure, and its graph-based analogues have recently gained prominence as practical tools for reweighting, pruning, and reshaping network geometry. We propose Sobolev-Ricci Curvature (SRC), a graph Ricci curvature canonically induced by Sobolev transport geometry, which admits efficient evaluation via a tree-metric Sobolev structure on neighborhood measures. We establish two consistency behaviors that anchor SRC to classical transport curvature: (i) on trees endowed with the length measure, SRC recovers Ollivier-Ricci curvature (ORC) in the canonical W1 setting, and (ii) SRC vanishes in the Dirac limit, matching the flat case of measure-theoretic Ricci curvature. We demonstrate SRC as a reusable curvature primitive in two representative pipelines. We define Sobolev-Ricci Flow by replacing ORC with SRC in a Ricci-flow-style reweighting rule, and we use SRC for curvature-guided edge pruning aimed at preserving manifold structure. Overall, SRC provides a transport-based foundation for scalable curvature-driven graph transformation and manifold-oriented pruning.

Method: Clients generate micro-subclusters and upload prototypes to server for hierarchical density-based merging. Progressive merging self-terminates based on neighboring relationships to determine optimal $k^*$.

Result: Extensive experiments on diverse datasets demonstrate accurate exploration of proper cluster numbers, handling varying cluster sizes and shapes in federated settings.

Conclusion: Fed-$k^*$-HC effectively addresses key challenges in federated clustering by automatically determining optimal cluster count while preserving privacy and handling real-world data distributions.

Abstract: Federated Clustering (FC) is an emerging and promising solution in exploring data distribution patterns from distributed and privacy-protected data in an unsupervised manner. Existing FC methods implicitly rely on the assumption that clients are with a known number of uniformly sized clusters. However, the true number of clusters is typically unknown, and cluster sizes are naturally imbalanced in real scenarios. Furthermore, the privacy-preserving transmission constraints in federated learning inevitably reduce usable information, making the development of robust and accurate FC extremely challenging. Accordingly, we propose a novel FC framework named Fed-$k^$-HC, which can automatically determine an optimal number of clusters $k^$ based on the data distribution explored through hierarchical clustering. To obtain the global data distribution for $k^$ determination, we let each client generate micro-subclusters. Their prototypes are then uploaded to the server for hierarchical merging. The density-based merging design allows exploring clusters of varying sizes and shapes, and the progressive merging process can self-terminate according to the neighboring relationships among the prototypes to determine $k^$. Extensive experiments on diverse datasets demonstrate the FC capability of the proposed Fed-$k^*$-HC in accurately exploring a proper number of clusters.

Method: Proposes Disentangled Latent Dynamics Manifold Fusion (DLDMF) that explicitly separates space, time, and parameters. Maps PDE parameters directly to continuous latent embeddings via feed-forward network, initializes latent state conditioned on these embeddings, and evolves it using parameter-conditioned Neural ODE. Uses dynamic manifold fusion with shared decoder to combine spatial coordinates, parameter embeddings, and time-evolving latent states to reconstruct spatiotemporal solutions.

Result: DLDMF consistently outperforms state-of-the-art baselines in accuracy, parameter generalization, and extrapolation robustness across several benchmark problems. Successfully handles unseen parameter settings and long-term temporal extrapolation.

Conclusion: By modeling prediction as latent dynamic evolution rather than static coordinate fitting, DLDMF reduces interference between parameter variation and temporal evolution while preserving smooth, coherent solution manifolds, enabling effective parameter generalization and temporal extrapolation.

Abstract: Generalizing neural surrogate models across different PDE parameters remains difficult because changes in PDE coefficients often make learning harder and optimization less stable. The problem becomes even more severe when the model must also predict beyond the training time range. Existing methods usually cannot handle parameter generalization and temporal extrapolation at the same time. Standard parameterized models treat time as just another input and therefore fail to capture intrinsic dynamics, while recent continuous-time latent methods often rely on expensive test-time auto-decoding for each instance, which is inefficient and can disrupt continuity across the parameterized solution space. To address this, we propose Disentangled Latent Dynamics Manifold Fusion (DLDMF), a physics-informed framework that explicitly separates space, time, and parameters. Instead of unstable auto-decoding, DLDMF maps PDE parameters directly to a continuous latent embedding through a feed-forward network. This embedding initializes and conditions a latent state whose evolution is governed by a parameter-conditioned Neural ODE. We further introduce a dynamic manifold fusion mechanism that uses a shared decoder to combine spatial coordinates, parameter embeddings, and time-evolving latent states to reconstruct the corresponding spatiotemporal solution. By modeling prediction as latent dynamic evolution rather than static coordinate fitting, DLDMF reduces interference between parameter variation and temporal evolution while preserving a smooth and coherent solution manifold. As a result, it performs well on unseen parameter settings and in long-term temporal extrapolation. Experiments on several benchmark problems show that DLDMF consistently outperforms state-of-the-art baselines in accuracy, parameter generalization, and extrapolation robustness.

Method: Transformer-based ensemble framework integrating protein language modeling and active learning to capture both high-level sequence semantics and fine-grained transformation patterns, enabling direct reaction prediction without EC numbers.

Result: Consistent improvements over six EC-based baselines with gains of 16.54% in F1 score and 15.43% in accuracy on curated cross-validation and temporal test sets.

Conclusion: RXNRECer provides a robust EC-free solution for fine-grained enzyme function prediction with advantages for proteome-wide annotation, reaction schema refinement, uncurated protein annotation, and enzyme promiscuity identification.

Abstract: A key challenge in enzyme annotation is identifying the biochemical reactions catalyzed by proteins. Most existing methods rely on Enzyme Commission (EC) numbers as intermediaries: they first predict an EC number and then retrieve the associated reactions. This indirect strategy introduces ambiguity due to the complex many-to-many mappings among proteins, EC numbers, and reactions, and is further complicated by frequent updates to EC numbers and inconsistencies across databases. To address these challenges, we present RXNRECer, a transformer-based ensemble framework that directly predicts enzyme-catalyzed reactions without relying on EC numbers. It integrates protein language modeling and active learning to capture both high-level sequence semantics and fine-grained transformation patterns. Evaluations on curated cross-validation and temporal test sets demonstrate consistent improvements over six EC-based baselines, with gains of 16.54% in F1 score and 15.43% in accuracy. Beyond accuracy gains, the framework offers clear advantages for downstream applications, including scalable proteome-wide reaction annotation, enhanced specificity in refining generic reaction schemas, systematic annotation of previously uncurated proteins, and reliable identification of enzyme promiscuity. By incorporating large language models, it also provides interpretable rationales for predictions. These capabilities make RXNRECer a robust and versatile solution for EC-free, fine-grained enzyme function prediction, with potential applications across multiple areas of enzyme research and industrial applications.

Method: Proposes modality-level partitioning at the boundary between vision encoder and language model, which minimizes cross-device transfer. Develops a closed-form cost model and builds HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling.

Result: Modality-level partitioning reduces transfer complexity from O(Ls_ctx) bytes to O(N_vd) bytes (O(L) reduction). HeteroServe improves throughput by up to 54% on identical hardware and achieves 37% better Tokens/$ with 40.6% cost savings in heterogeneous deployment.

Conclusion: Modality-level disaggregation enables efficient cross-tier heterogeneous serving over commodity PCIe, providing significant cost savings and performance improvements for multimodal LLM deployment.

Abstract: Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster ($38k) improves Tokens/$ by 37% over a homogeneous baseline ($64k) without degrading latency.

Method: Created SciDesignBench with 520 simulator-grounded tasks across 14 scientific domains and 5 settings (single-shot design, short-horizon feedback, long-horizon refinement, seed-design optimization). Evaluated state-of-the-art models, introduced RLSF (simulator-feedback training recipe) to improve model performance through training on simulator feedback.

Result: Best zero-shot model achieved only 29.0% success on 10-domain subset. Simulator feedback helps but different models excel at different horizons: Sonnet 4.5 strongest in one-turn de novo design, Opus 4.6 strongest after 20 turns of refinement. RLSF-tuned 8B model improved single-turn success rates by 8-17 percentage points across three domains.

Conclusion: Simulator-grounded inverse design serves as both a benchmark for scientific reasoning and a practical approach for amortizing expensive test-time compute into model weights. Different design settings require fundamentally different capabilities, and targeted training can significantly improve performance.

Abstract: Many of the most important problems in science and engineering are inverse problems: given a desired outcome, find a design that achieves it. Evaluating whether a candidate meets the spec is often routine; a binding energy can be computed, a reactor yield simulated, a pharmacokinetic profile predicted. But searching a combinatorial design space for inputs that satisfy those targets is fundamentally harder. We introduce SciDesignBench, a benchmark of 520 simulator-grounded tasks across 14 scientific domains and five settings spanning single-shot design, short-horizon feedback, long-horizon refinement, and seed-design optimization. On the 10-domain shared-core subset, the best zero-shot model reaches only 29.0% success despite substantially higher parse rates. Simulator feedback helps, but the leaderboard changes with horizon: Sonnet 4.5 is strongest in one-turn de novo design, whereas Opus 4.6 is strongest after 20 turns of simulator-grounded refinement. Providing a starting seed design reshuffles the leaderboard again, demonstrating that constrained modification requires a fundamentally different capability from unconstrained de novo generation. We then introduce RLSF, a simulator-feedback training recipe. An RLSF-tuned 8B model raises single-turn success rates by 8-17 percentage points across three domains. Together, these results position simulator-grounded inverse design as both a benchmark for scientific reasoning and a practical substrate for amortizing expensive test-time compute into model weights.

Method: Propose GICON (Graph In-Context Operator Network) combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Conduct controlled experiments comparing in-context vs classical operator learning on air quality prediction across Chinese regions.

Result: In-context operator learning outperforms classical operator learning on complex tasks, generalizes better across spatial domains, and scales robustly from few training examples to 100 at inference.

Conclusion: In-context operator learning provides advantages over single-operator learning even with identical training data, demonstrating better generalization and scalability for real-world spatiotemporal systems.

Abstract: In-context operator learning enables neural networks to infer solution operators from contextual examples without weight updates. While prior work has demonstrated the effectiveness of this paradigm in leveraging vast datasets, a systematic comparison against single-operator learning using identical training data has been absent. We address this gap through controlled experiments comparing in-context operator learning against classical operator learning (single-operator models trained without contextual examples), under the same training steps and dataset. To enable this investigation on real-world spatiotemporal systems, we propose GICON (Graph In-Context Operator Network), combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Experiments on air quality prediction across two Chinese regions show that in-context operator learning outperforms classical operator learning on complex tasks, generalizing across spatial domains and scaling robustly from few training examples to 100 at inference.

Method: Created TaoBench from Terence Tao’s Analysis I, formalizing analysis concepts from scratch without MathLib. Built an agentic pipeline for compilable, self-contained environments. Translated problems into mathematically equivalent MathLib formulations for paired comparison.

Result: State-of-the-art ATP models perform capably within MathLib framework but performance drops by ~26% on definitionally equivalent Tao formulations, indicating limited generalization across definitional frameworks rather than task difficulty.

Conclusion: TaoBench highlights a gap between benchmark performance and real-world applicability, providing a foundation for developing provers better aligned with research mathematics that can handle diverse definitional frameworks.

Abstract: Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib’s definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao’s Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.

Method: Develops an upper-bound formula for local learning coefficients at singular points in three-layer neural networks. The formula can be interpreted as a counting rule under budget constraints and demand-supply constraints, and is applicable to general analytic activation functions.

Result: When input dimension is one, the derived upper bound coincides with already known learning coefficients, partially resolving previous discrepancies. The result provides systematic understanding of how weight parameters affect learning coefficients in three-layer networks.

Conclusion: The paper extends previous results to wider class of activation functions and provides a formula applicable at singular points, offering better theoretical understanding of learning coefficients in three-layer neural networks.

Abstract: Three-layer neural networks are known to form singular learning models, and their Bayesian asymptotic behavior is governed by the learning coefficient, or real log canonical threshold. Although this quantity has been clarified for regular models and for some special singular models, broadly applicable methods for evaluating it in neural networks remain limited. Recently, a formula for the local learning coefficient of semiregular models was proposed, yielding an upper bound on the learning coefficient. However, this formula applies only to nonsingular points in the set of realization parameters and cannot be used at singular points. In particular, for three-layer neural networks, the resulting upper bound has been shown to differ substantially from learning coefficient values already known in some cases. In this paper, we derive an upper-bound formula for the local learning coefficient at singular points in three-layer neural networks. This formula can be interpreted as a counting rule under budget constraints and demand-supply constraints, and is applicable to general analytic activation functions. In particular, it covers the swish function and polynomial functions, extending previous results to a wider class of activation functions. We further show that, when the input dimension is one, the upper bound obtained here coincides with the already known learning coefficient, thereby partially resolving the discrepancy above. Our result also provides a systematic perspective on how the weight parameters of three-layer neural networks affect the learning coefficient.

Method: Proposes a novel class of non-stationary kernels derived from the fundamental solution of generalized time-space fractional diffusion-wave equations. Uses structure-preserving transmutation method over Weighted Sobolev Spaces to create the Fox-Dorrego Kernel, an exact analytical Mercer kernel governed by the Fox H-function. Incorporates an aging weight function (“Amnesia Effect”) to penalize distant outliers and fractional asymptotic power-law decay for robust, heavy-tailed feature mapping.

Result: Numerical experiments on synthetic datasets and real-world high-dimensional radar data (Ionosphere) show the Fox-Dorrego kernel consistently outperforms standard Gaussian RBF baseline, reducing classification error rate by approximately 50% while maintaining structural robustness against outliers.

Conclusion: The proposed Fox-Dorrego kernel provides a superior alternative to Gaussian RBF kernels, offering better classification performance and enhanced robustness to outliers through its fractional diffusion-based formulation and amnesia effect mechanism.

Abstract: Support Vector Machines (SVMs) rely heavily on the choice of the kernel function to map data into high-dimensional feature spaces. While the Gaussian Radial Basis Function (RBF) is the industry standard, its exponential decay makes it highly susceptible to structural noise and outliers, often leading to severe overfitting in complex datasets. In this paper, we propose a novel class of non-stationary kernels derived from the fundamental solution of the generalized time-space fractional diffusion-wave equation. By leveraging a structure-preserving transmutation method over Weighted Sobolev Spaces, we introduce the Fox-Dorrego Kernel, an exact analytical Mercer kernel governed by the Fox H-function. Unlike standard kernels, our formulation incorporates an aging weight function (the “Amnesia Effect”) to penalize distant outliers and a fractional asymptotic power-law decay to allow for robust, heavy-tailed feature mapping (analogous to Lévy flights). Numerical experiments on both synthetic datasets and real-world high-dimensional radar data (Ionosphere) demonstrate that the proposed Fox-Dorrego kernel consistently outperforms the standard Gaussian RBF baseline, reducing the classification error rate by approximately 50% while maintaining structural robustness against outliers.

Method: Multi-task large reasoning model with multi-specialist modules for molecular expertise, chain-of-thought framework enhanced by reinforcement learning infused with molecular knowledge, enabling structured and reflective reasoning.

Result: Achieved 50.3% average improvement over base architecture across 10 molecular tasks and 47 metrics, outperforming 20+ state-of-the-art baselines including ultra-large foundation models despite using significantly fewer training data and computational resources.

Conclusion: Embedding explicit reasoning mechanisms enables high-efficiency learning, allowing smaller models to surpass massive counterparts in efficacy and interpretability. The framework bridges data-driven and knowledge-integrated approaches for intelligent molecular design.

Abstract: Advancements in artificial intelligence for molecular science are necessitating a paradigm shift from purely data-driven predictions to knowledge-guided computational reasoning. Existing molecular models are predominantly proprietary, lacking general molecular intelligence and generalizability. This underscores the necessity for computational methods that can effectively integrate scientific logic with deep learning architectures. Here we introduce a multi-task large reasoning model designed to emulate the cognitive processes of molecular scientists through structured reasoning and reflection. Our approach incorporates multi-specialist modules to provide versatile molecular expertise and a chain-of-thought (CoT) framework enhanced by reinforcement learning infused with molecular knowledge, enabling structured and reflective reasoning. Systematic evaluations across 10 molecular tasks and 47 metrics demonstrate that our model achieves an average 50.3% improvement over the base architecture, outperforming over 20 state-of-the-art baselines, including ultra-large-parameter foundation models, despite using significantly fewer training data and computational resources. This validates that embedding explicit reasoning mechanisms enables high-efficiency learning, allowing smaller-scale models to surpass massive counterparts in both efficacy and interpretability. The practical utility of this computational framework was validated through a case study on the design of central nervous system (CNS) drug candidates, illustrating its capacity to bridge data-driven and knowledge-integrated approaches for intelligent molecular design.

Method: Proposes Residual SODAP framework that jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Combines α-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage-based drift detection, and uncertainty-aware multi-loss balancing.

Result: Achieves state-of-the-art performance across three DIL benchmarks without task IDs or extra data storage: AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).

Conclusion: Residual SODAP effectively addresses catastrophic forgetting in domain-incremental learning by combining representation adaptation with classifier preservation, demonstrating strong performance across diverse benchmarks.

Abstract: Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $α$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage–based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).

Method: Proposes a novel outlier detection paradigm using graph structures to leverage natural neighboring relationships. Incorporates reference sets at both local and global scales derived from the graph for multi-perspective anomaly evaluation. The approach recognizes scattered outliers without interference from clustered anomalies while using graph structure to reflect and isolate clustered outlier groups.

Result: Extensive experiments including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and hyperparameter sensitivity evaluation demonstrate the method’s efficacy. The source code is publicly available.

Conclusion: The graph-based approach effectively addresses both scattered and clustered outliers in IoT data analysis, overcoming limitations of traditional methods that struggle with clustered anomalies being mistaken for normal behavior.

Abstract: Most real-world IoT data analysis tasks, such as clustering and anomaly event detection, are unsupervised and highly susceptible to the presence of outliers. In addition to sporadic scattered outliers caused by factors such as faulty sensor readings, IoT systems often exhibit clustered outliers. These occur when multiple devices or nodes produce similar anomalous measurements, for instance, owing to localized interference, emerging security threats, or regional false alarms, forming micro-clusters. These clustered outliers can be easily mistaken for normal behavior because of their relatively high local density, thereby obscuring the detection of both scattered and contextual anomalies. To address this, we propose a novel outlier detection paradigm that leverages the natural neighboring relationships using graph structures. This facilitates multi-perspective anomaly evaluation by incorporating reference sets at both local and global scales derived from the graph. Our approach enables the effective recognition of scattered outliers without interference from clustered anomalies, whereas the graph structure simultaneously helps reflect and isolate clustered outlier groups. Extensive experiments, including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and evaluation of hyperparameter sensitivity, demonstrate the efficacy of the proposed method. The source code is available at https://github.com/gordonlok/DROD.

Method: The paper reviews theoretical approaches to assessing linear separability and state-of-the-art methods/tools, then systematically examines all relevant assemblies of the MNIST dataset. It distinguishes between pairwise separation and one-vs-rest separation for training, test, and combined sets respectively.

Result: The paper provides comprehensive empirical findings about the linear separability of MNIST dataset, addressing the long-standing question with systematic analysis across different dataset partitions and separation approaches.

Conclusion: The study offers definitive empirical evidence about the linear separability properties of MNIST, resolving conflicting claims and providing a comprehensive understanding of this fundamental dataset’s characteristics.

Abstract: The MNIST dataset containing thousands of handwritten digit images is still a fundamental benchmark for evaluating various pattern-recognition and image-classification models. Linear separability is a key concept in many statistical and machine-learning techniques. Despite the long history of the MNIST dataset and its relative simplicity in size and resolution, the question of whether the dataset is linearly separable has never been fully answered – scientific and informal sources share conflicting claims. This paper aims to provide a comprehensive empirical investigation to address this question, distinguishing pairwise and one-vs-rest separation of the training, the test and the combined sets, respectively. It reviews the theoretical approaches to assessing linear separability, alongside state-of-the-art methods and tools, then systematically examines all relevant assemblies, and reports the findings.

Method: Two-stage test-time RL alignment: 1) RL with a single sample aligns model to task format, 2) test-time RL with majority-voting reward aligns model to benchmark distribution. This eliminates need for task-specific training data while achieving similar alignment as SFT-based train-before-test.

Result: On domain-specific benchmarks without training data, direct evaluation underestimates base models which perform substantially better once aligned. For reasoning tasks, performance gap between fine-tuned models and their base models largely disappears after alignment, suggesting many reported RL/SFT gains are artifacts of task familiarity rather than reasoning capability improvements.

Conclusion: Test-time RL alignment provides a more faithful evaluation of LLM capabilities by controlling for task familiarity without requiring task-specific training data. Many reported gains from RL/SFT in literature may not reflect actual reasoning capability improvements but rather artifacts of task familiarity.

Abstract: Direct evaluation of LLMs on benchmarks can be misleading because comparatively strong performance may reflect task familiarity rather than capability. The train-before-test approach controls for task familiarity by giving each model task-relevant training before evaluation, originally through supervised finetuning. However, suitable training data is often hard to come by, and evaluation results vary with the data chosen. In this paper, we propose a two-stage test-time reinforcement learning (RL) alignment method for train-before-test. First, RL with a single sample provides a first alignment of the model to the task format, and second, test-time RL with majority-voting reward aligns the model to the benchmark distribution. Our test-time RL alignment method aligns similarly well as SFT-based train-before test, but without requiring a task-specific training set. On a domain-specific benchmark without training data, we show that direct evaluation underestimates base models which perform substantially better once aligned, yielding a more faithful evaluation of their capabilities. Moreover, for reasoning tasks, the performance gap between fine-tuned models and their base models largely disappears after alignment, suggesting that many gains from RLVR/SFT reported in the literature are not a difference in reasoning capability, but rather artifacts of task familiarity.

Method: Proposes a knowledge augmentation framework using reinforcement learning techniques to adaptively extract and synthesize prior drug knowledge, optimizing the strategy space to enhance LLM accuracy for DDIE predictions through few-shot learning.

Result: Achieved notable improvement compared to baseline methods through few-shot learning, establishing an effective framework for scientific knowledge learning in DDIE predictions.

Conclusion: The reinforcement learning-based knowledge augmentation framework successfully enhances LLM performance for DDIE prediction, addressing key challenges in the field and providing an effective approach for scientific knowledge integration.

Abstract: Drug-drug interaction event (DDIE) prediction is crucial for preventing adverse reactions and ensuring optimal therapeutic outcomes. However, existing methods often face challenges with imbalanced datasets, complex interaction mechanisms, and poor generalization to unknown drug combinations. To address these challenges, we propose a knowledge augmentation framework that adaptively infuses prior drug knowledge into a large language model (LLM). This framework utilizes reinforcement learning techniques to facilitate adaptive knowledge extraction and synthesis, thereby efficiently optimizing the strategy space to enhance the accuracy of LLMs for DDIE predictions. As a result of few-shot learning, we achieved a notable improvement compared to the baseline. This approach establishes an effective framework for scientific knowledge learning for DDIE predictions.

Method: Extends Dirichlet Prior Augmentation (DirPA) method to proactively mitigate label distribution skews during training. Evaluates across multiple EU countries to test resilience across diverse agricultural environments beyond localized experiments.

Result: DirPA demonstrates effectiveness across different geographical regions, improving system robustness and stabilizing training under extreme long-tailed distributions regardless of target region. Substantially improves individual class-specific performance by proactively simulating priors.

Conclusion: DirPA successfully addresses distribution shift challenges in agricultural few-shot learning across diverse European environments, making models more robust to real-world class imbalances.

Abstract: Real-world agricultural monitoring is often hampered by severe class imbalance and high label acquisition costs, resulting in significant data scarcity. In few-shot learning (FSL) – a framework specifically designed for data-scarce settings – , training sets are often artificially balanced. However, this creates a disconnect from the long-tailed distributions observed in nature, leading to a distribution shift that undermines the model’s ability to generalize to real-world agricultural tasks. We previously introduced Dirichlet Prior Augmentation (DirPA; Reuss et al., 2026a) to proactively mitigate the effects of such label distribution skews during model training. In this work, we extend the original study’s geographical scope. Specifically, we evaluate this extended approach across multiple countries in the European Union (EU), moving beyond localized experiments to test the method’s resilience across diverse agricultural environments. Our results demonstrate the effectiveness of DirPA across different geographical regions. We show that DirPA not only improves system robustness and stabilizes training under extreme long-tailed distributions, regardless of the target region, but also substantially improves individual class-specific performance by proactively simulating priors.

Method: Uses multi-head attention query evolution as a predictable process with dual pathways: gradient-updated reconstruction and history-only predictor that forecasts future query vectors from past context. Trained via masked predictor-target objective against EMA target encoder. Combines reconstruction error with tail-aggregated query mismatch score at inference.

Result: Outperforms strong baselines on proprietary in-vehicle telemetry and TSB-AD multi-variate suite (17 datasets, 180 series) using threshold-free and range-aware metrics, improving ranking quality and temporal localization.

Conclusion: AxonAD effectively detects structural dependency shifts in multivariate time series while retaining amplitude-level detection, with query prediction and combined scoring being key to performance gains.

Abstract: Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL https://github.com/iis-esslingen/AxonAD.

Method: SCOPE analyzes latent space distribution using three scores: representation score (reliability of core class features), diversity score (novelty of orthogonal residuals), and boundary proximity score (similarity to competing classes). It shares only scalar metrics with a federated server to build global consensus.

Result: SCOPE achieves competitive global accuracy and robust convergence with 128x-512x reduction in uplink bandwidth, 7.72x wall-clock acceleration, and reduced FLOP/VRAM footprints for local coreset selection.

Conclusion: SCOPE provides an efficient federated learning framework that addresses class imbalance through semantic coreset selection while maintaining communication efficiency and performance.

Abstract: Scientific discovery increasingly requires learning on federated datasets, fed by streams from high-resolution instruments, that have extreme class imbalance. Current ML approaches either require impractical data aggregation or fail due to class imbalance. Existing coreset selection methods rely on local heuristics, making them unaware of the global data landscape and prone to sub-optimal and non-representative pruning. To overcome these challenges, we introduce SCOPE (Semantic Coreset using Orthogonal Projection Embeddings for Federated learning), a coreset framework for federated data that filters anomalies and adaptively prunes redundant data to mitigate long-tail skew. By analyzing the latent space distribution, we score each data point using a representation score that measures the reliability of core class features, a diversity score that quantifies the novelty of orthogonal residuals, and a boundary proximity score that indicates similarity to competing classes. Unlike prior methods, SCOPE shares only scalar metrics with a federated server to construct a global consensus, ensuring communication efficiency. Guided by the global consensus, SCOPE dynamically filters local noise and discards redundant samples to counteract global feature skews. Extensive experiments demonstrate that SCOPE yields competitive global accuracy and robust convergence, all while achieving exceptional efficiency with a 128x to 512x reduction in uplink bandwidth, a 7.72x wall-clock acceleration and reduced FLOP and VRAM footprints for local coreset selection.

Method: Uses frozen foundation model with ridge-regression head where exact optimum depends on data through two additive sufficient statistics; creates communication protocol supporting add/delete requests via fixed-size messages; server maintains head pointwise identical to centralized retraining

Result: Method matches centralized ridge retraining to within 10^-9 relative Frobenius error; provides deterministic retrain-equivalence guarantees, order/partition invariance, and Bayesian certificate of zero KL divergence; completes requests at orders-of-magnitude faster than retraining

Conclusion: Proposes efficient federated unlearning method for frozen foundation models that provides exact retraining equivalence with strong theoretical guarantees and practical efficiency

Abstract: Foundation models are commonly deployed as frozen feature extractors with a small trainable head to adapt to private, user-generated data in federated settings. The ``right to be forgotten’’ requires removing the influence of specific samples or users from the trained model on demand. Existing federated unlearning methods target general deep models and rely on approximate reconstruction or selective retraining, making exactness costly or elusive. We study this problem in a practically relevant but under-explored regime: a frozen foundation model with a ridge-regression head. The exact optimum depends on the data only through two additive sufficient statistics, which we turn into a communication protocol supporting an arbitrary stream of \emph{add} and \emph{delete} requests via fixed-size messages. The server maintains a head that is, in exact arithmetic, \emph{pointwise identical} to centralized retraining after every request. We provide deterministic retrain-equivalence guarantees, order and partition invariance, two server-side variants, and a Bayesian certificate of zero KL divergence. Experiments on four benchmarks confirm the guarantees: both variants match centralized ridge retraining to within $10^{-9}$ relative Frobenius error and complete each request at orders-of-

Method: Proposes a hybrid vector-geographical retrieval module that learns to select comparables, jointly optimized with an estimation module. The system adapts to different datasets and learns selection policies rather than imposing fixed rules.

Result: Demonstrates that learned selection policies significantly improve comparable selection over state-of-the-art methods. Shows that carefully selected comparables enable models with fewer comparables and parameters while maintaining performance close to state-of-the-art.

Conclusion: Learning comparable selection policies is more effective than imposing fixed rules for real estate appraisal. The proposed hybrid approach adapts well across diverse geographical datasets (US, Brazil, France) and enables more efficient models.

Abstract: The Sales Comparison Approach (SCA) is one of the most popular when it comes to real estate appraisal. Used as a reference in real estate expertise and as one of the major types of Automatic Valuation Models (AVM), it recently gained popularity within machine learning methods. The performance of models able to use data represented as sets and graphs made it possible to adapt this methodology efficiently, yielding substantial results. SCA relies on taking past transactions (comparables) as references, selected according to their similarity with the target property’s sale. In this study, we focus on the selection of these comparables for real estate appraisal. We demonstrate that the selection of comparables used in many state-of-the-art algorithms can be significantly improved by learning a selection policy instead of imposing it. Our method relies on a hybrid vector-geographical retrieval module capable of adapting to different datasets and optimized jointly with an estimation module. We further show that the use of carefully selected comparables makes it possible to build models that require fewer comparables and fewer parameters with performance close to state-of-the-art models. All our evaluations are made on five datasets which span areas in the United States, Brazil, and France.

Method: DAPD uses self-attention to induce a conditional dependency graph over masked tokens, where edges capture strong token interactions and non-edges indicate weak dependence. It then selects an independent set on the graph to unmask tokens in parallel without co-updating strongly coupled tokens.

Result: Experiments on LLaDA and Dream show DAPD improves the accuracy-steps trade-off over existing methods and enables more globally distributed parallel updates that better exploit the any-order generation capability of diffusion LLMs.

Conclusion: DAPD provides a simple, training-free decoding method that effectively addresses the parallel decoding challenge in diffusion LLMs by leveraging attention-based dependency graphs.

Abstract: Parallel decoding for diffusion LLMs (dLLMs) is difficult because each denoising step provides only token-wise marginal distributions, while unmasking multiple tokens simultaneously requires accounting for inter-token dependencies. We propose Dependency-Aware Parallel Decoding (DAPD), a simple, training-free decoding method that uses self-attention to induce a conditional dependency graph over masked tokens. At each iteration, edges in this graph capture strong token interactions, while non-edges indicate weak dependence. Parallel decoding is then reduced to selecting an independent set on the graph and unmasking the selected tokens in parallel. This avoids co-updating strongly coupled tokens without auxiliary models or retraining. Experiments on LLaDA and Dream show that DAPD improves the accuracy-steps trade-off over existing methods and enables more globally distributed parallel updates that better exploit the any-order generation capability of dLLMs.

Method: Conducted experiments under idealized conditions providing correction methods with a perfect oracle transition matrix, analyzed macroscopic convergence states, microscopic optimization dynamics, and information-theoretic limits.

Result: Even with perfect transition matrix, noise-correction methods still suffer from performance collapse during training, demonstrating failure is not fundamentally a T-estimation problem but stems from deeper flaws.

Conclusion: The failure of ideal noise correction reveals fundamental limitations in current theoretical approaches to learning with noisy labels, providing guidance for designing more reliable methods.

Abstract: Statistically consistent methods based on the noise transition matrix ($T$) offer a theoretically grounded solution to Learning with Noisy Labels (LNL), with guarantees of convergence to the optimal clean-data classifier. In practice, however, these methods are often outperformed by empirical approaches such as sample selection, and this gap is usually attributed to the difficulty of accurately estimating $T$. The common assumption is that, given a perfect $T$, noise-correction methods would recover their theoretical advantage. In this work, we put this longstanding hypothesis to a decisive test. We conduct experiments under idealized conditions, providing correction methods with a perfect, oracle transition matrix. Even under these ideal conditions, we observe that these methods still suffer from performance collapse during training. This compellingly demonstrates that the failure is not fundamentally a $T$-estimation problem, but stems from a more deeply rooted flaw. To explain this behaviour, we provide a unified analysis that links three levels: macroscopic convergence states, microscopic optimisation dynamics, and information-theoretic limits on what can be learned from noisy labels. Together, these results give a formal account of why ideal noise correction fails and offer concrete guidance for designing more reliable methods for learning with noisy labels.

Method: Reinforcement learning framework that trains an attack LLM to optimize injected prompts in black-box setting. Introduces adaptive entropy regularization to sustain exploration and dynamic advantage weighting to amplify learning from scarce successes.

Result: State-of-the-art prompt injection defenses remain vulnerable to adaptive attacks. PISmith outperforms 7 baselines across static, search-based, and RL-based attack categories, achieving highest attack success rates on 13 benchmarks and strong performance in agentic settings.

Conclusion: Current prompt injection defenses are insufficient against sophisticated adaptive attacks. PISmith provides a systematic evaluation framework that reveals vulnerabilities and highlights the need for more robust defense mechanisms.

Abstract: Prompt injection poses serious security risks to real-world LLM applications, particularly autonomous agents. Although many defenses have been proposed, their robustness against adaptive attacks remains insufficiently evaluated, potentially creating a false sense of security. In this work, we propose PISmith, a reinforcement learning (RL)-based red-teaming framework that systematically assesses existing prompt-injection defenses by training an attack LLM to optimize injected prompts in a practical black-box setting, where the attacker can only query the defended LLM and observe its outputs. We find that directly applying standard GRPO to attack strong defenses leads to sub-optimal performance due to extreme reward sparsity – most generated injected prompts are blocked by the defense, causing the policy’s entropy to collapse before discovering effective attack strategies, while the rare successes cannot be learned effectively. In response, we introduce adaptive entropy regularization and dynamic advantage weighting to sustain exploration and amplify learning from scarce successes. Extensive evaluation on 13 benchmarks demonstrates that state-of-the-art prompt injection defenses remain vulnerable to adaptive attacks. We also compare PISmith with 7 baselines across static, search-based, and RL-based attack categories, showing that PISmith consistently achieves the highest attack success rates. Furthermore, PISmith achieves strong performance in agentic settings on InjecAgent and AgentDojo against both open-source and closed-source LLMs (e.g., GPT-4o-mini and GPT-5-nano). Our code is available at https://github.com/albert-y1n/PISmith.

Method: Two-level optimization approach: (1) accuracy-constrained architecture search using GNN-based surrogate models for PPA and error prediction, exploring compressor combinations and SRAM macro parameters; (2) variation-aware transistor sizing for standard cells and SRAM bitcells using Monte Carlo methods, decoupled for efficient optimization.

Result: Significant PPA improvements under controlled accuracy budgets, enabling rapid “what-if” exploration for approximate DCiM designs, with the framework being compatible with FreePDK45 and OpenROAD for reproducible evaluation.

Conclusion: OpenACMv2 provides an effective framework for accuracy-constrained co-optimization of approximate DCiM architectures, delivering strong PPA-accuracy tradeoffs and robust convergence through decoupled two-level optimization.

Abstract: Digital Compute-in-Memory (DCiM) accelerates neural networks by reducing data movement. Approximate DCiM can further improve power-performance-area (PPA), but demands accuracy-constrained co-optimization across coupled architecture and transistor-level choices. Building on OpenYield, we introduce Accuracy-Constrained Co-Optimization (ACCO) and present OpenACMv2, an open framework that operationalizes ACCO via two-level optimization: (1) accuracy-constrained architecture search of compressor combinations and SRAM macro parameters, driven by a fast GNN-based surrogate for PPA and error; and (2) variation- and PVT-aware transistor sizing for standard cells and SRAM bitcells using Monte Carlo. By decoupling ACCO into architecture-level exploration and circuit-level sizing, OpenACMv2 integrates classic single- and multi-objective optimizers to deliver strong PPA-accuracy tradeoffs and robust convergence. The workflow is compatible with FreePDK45 and OpenROAD, supporting reproducible evaluation and easy adoption. Experiments demonstrate significant PPA improvements under controlled accuracy budgets, enabling rapid “what-if” exploration for approximate DCiM. The framework is available on https://github.com/ShenShan123/OpenACM.

Method: Physics-based generative framework using conditional Flow Matching (CFM) trained on 6-year, 3-km-resolution WRF dataset. Employs latent domain adaptation and two-stage transfer learning for region-adaptive vortex-following reconstruction.

Result: Outperforms ECMWF-HRES in TC intensity prediction at nearly all lead times up to 5 days, reduces RMSE of maximum WS10M by 36.5% relative to FuXi inputs, improves representation of TC inner-core structure while maintaining track stability.

Conclusion: 3DTCR offers an efficient physics-based generative approach for resolving fine-scale TC structures at lower computational cost, providing a promising avenue for improving TC intensity forecasting.

Abstract: Tropical cyclone (TC) intensity forecasting remains challenging as current numerical and AI-based weather models fail to satisfactorily represent extreme TC structure and intensity. Although intensity time-series forecasting has achieved significant advances, it outputs intensity sequences rather than the three-dimensional inner-core fine-scale structure and physical mechanisms governing TC evolution. High-resolution numerical simulations can capture these features but remain computationally expensive and inefficient for large-scale operational applications. Here we present 3DTCR, a physics-based generative framework combining physical constraints with generative AI efficiency for 3D TC structure reconstruction. Trained on a six-year, 3-km-resolution moving-domain WRF dataset, 3DTCR enables region-adaptive vortex-following reconstruction using conditional Flow Matching(CFM), optimized via latent domain adaptation and two-stage transfer learning. The framework mitigates limitations imposed by low-resolution targets and over-smoothed forecasts, improving the representation of TC inner-core structure and intensity while maintaining track stability. Results demonstrate that 3DTCR outperforms the ECMWF high-resolution forecasting system (ECMWF-HRES) in TC intensity prediction at nearly all lead times up to 5 days and reduces the RMSE of maximum WS10M by 36.5% relative to its FuXi inputs. These findings highlight 3DTCR as a physics-based generative framework that efficiently resolves fine-scale structures at lower computational cost, which may offer a promising avenue for improving TC intensity forecasting.

Method: Leverages structured nature of biological interventions with specific inductive biases/invariances, uses available perturbation effect information to generalize to novel contexts, and requires only widely-available bulk molecular data with efficient architectures.

Result: Extensive testing shows accurate prediction of context-specific perturbation effects compared to real large-scale interventional experiments, competitive with state-of-the-art foundation models but with simpler data, smaller models, and less time.

Conclusion: Accurate perturbation effect prediction is possible without proprietary hardware or very large models by focusing on robust bulk signals and efficient architectures, opening ways to leverage causal learning in biomedicine.

Abstract: Predicting the effects of chemical and genetic perturbations on quantitative cell states is a central challenge in computational biology, molecular medicine and drug discovery. Recent work has leveraged large-scale single-cell data and massive foundation models to address this task. However, such computational resources and extensive datasets are not always accessible in academic or clinical settings, hence limiting utility. Here we propose a lightweight framework for perturbation effect prediction that exploits the structured nature of biological interventions and specific inductive biases/invariances. Our approach leverages available information concerning perturbation effects to allow generalization to novel contexts and requires only widely-available bulk molecular data. Extensive testing, comparing predictions of context-specific perturbation effects against real, large-scale interventional experiments, demonstrates accurate prediction in new contexts. The proposed approach is competitive with SOTA foundation models but requires simpler data, much smaller model sizes and less time. Focusing on robust bulk signals and efficient architectures, we show that accurate prediction of perturbation effects is possible without proprietary hardware or very large models, hence opening up ways to leverage causal learning approaches in biomedicine generally.

Method: Three complementary signals: (1) semantic neighborhoods from pretrained transformer representations of keyword text, (2) behavioral neighborhoods via Dynamic Time Warping alignment of CPC trajectories, and (3) geographic-intent covariates capturing localized demand heterogeneity. These are evaluated as stand-alone covariates and relational priors in spatiotemporal graph forecasters.

Result: Competition-aware augmentation improves forecast stability and error profiles at medium and longer horizons where competitive regime shifts are most consequential. The approach outperforms statistical, neural, and time-series foundation model baselines.

Conclusion: Broad market-outcome coverage combined with keyword-derived semantic and geographic priors provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.

Abstract: Cost-per-click (CPC) in paid search is a volatile auction outcome generated by a competitive landscape that is only partially observable from any single advertiser’s history. Using Google Ads auction logs from a concentrated car-rental market (2021–2023), we forecast weekly CPC for 1,811 keyword series and approximate latent competition through complementary signals derived from keyword text, CPC trajectories, and geographic market structure. We construct (i) semantic neighborhoods and a semantic keyword graph from pretrained transformer-based representations of keyword text, (ii) behavioral neighborhoods via Dynamic Time Warping (DTW) alignment of CPC trajectories, and (iii) geographic-intent covariates capturing localized demand and marketplace heterogeneity. We extensively evaluate these signals both as stand-alone covariates and as relational priors in spatiotemporal graph forecasters, benchmarking them against strong statistical, neural, and time-series foundation-model baselines. Across methods, competition-aware augmentation improves stability and error profiles at business-relevant medium and longer horizons, where competitive regimes shift and volatility is most consequential. The results show that broad market-outcome coverage, combined with keyword-derived semantic and geographic priors, provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.

Method: L2GTX aggregates local explanations from representative instances using parameterized temporal event primitives (trends, extrema). It clusters these events, merges across instances to reduce redundancy, uses importance matrices to estimate global relevance, and selects representative instances under a budget to maximize coverage of influential clusters.

Result: Experiments on six benchmark time series datasets show L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.

Conclusion: L2GTX provides an effective model-agnostic framework for generating class-wise global explanations for time series classification models, addressing limitations of existing XAI methods for temporal data.

Abstract: Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.

Method: Created GeoChemAD dataset from government geological surveys covering multiple regions, sampling sources, and target elements. Proposed GeoChemFormer, a transformer-based framework using self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples.

Result: GeoChemFormer consistently achieved superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability.

Conclusion: The dataset and framework provide a foundation for reproducible research and future development in geochemical anomaly detection, addressing limitations of previous studies.

Abstract: Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.

Method: The authors demonstrate that deterministic DDIM reverse chains operate as Partitioned Itered Function Systems (PIFS). They derive three computable geometric quantities from this structure: per-step contraction threshold, diagonal expansion function, and global expansion threshold. They analyze the fractal geometry of PIFS and use Kaplan-Yorke dimension and discrete Moran equation on Lyapunov spectrum.

Result: The PIFS framework explains diffusion model behavior: global context assembly at high noise via diffuse cross-patch attention and fine-detail synthesis at low noise via patch-by-patch suppression. The analysis reveals self-attention emerges as the natural primitive for PIFS contraction. The study derives three optimal design criteria that explain four prominent empirical design choices.

Conclusion: Diffusion models can be understood through geometric PIFS analysis, providing theoretical foundations that explain empirical design choices and offering computable metrics for analyzing denoising dynamics without model evaluation.

Abstract: What is a diffusion model actually doing when it turns noise into a photograph? We show that the deterministic DDIM reverse chain operates as a Partitioned Iterated Function System (PIFS) and that this framework serves as a unified design language for denoising diffusion model schedules, architectures, and training objectives. From the PIFS structure we derive three computable geometric quantities: a per-step contraction threshold $L^*_t$, a diagonal expansion function $f_t(λ)$ and a global expansion threshold $λ^{**}$. These quantities require no model evaluation and fully characterize the denoising dynamics. They structurally explain the two-regime behavior of diffusion models: global context assembly at high noise via diffuse cross-patch attention and fine-detail synthesis at low noise via patch-by-patch suppression release in strict variance order. Self-attention emerges as the natural primitive for PIFS contraction. The Kaplan-Yorke dimension of the PIFS attractor is determined analytically through a discrete Moran equation on the Lyapunov spectrum. Through the study of the fractal geometry of the PIFS, we derive three optimal design criteria and show that four prominent empirical design choices (the cosine schedule offset, resolution-dependent logSNR shift, Min-SNR loss weighting, and Align Your Steps sampling) each arise as approximate solutions to our explicit geometric optimization problems tuning theory into practice.

Method: Uses linearized attention mechanism with exact correspondence to a data-dependent Gram-induced kernel, analyzes through Neural Tangent Kernel (NTK) framework with both empirical and theoretical analysis, establishes spectral amplification result showing attention cubes Gram matrix condition number.

Result: Linearized attention does not converge to its infinite-width NTK limit even at large widths, requiring width m = Ω(κ^6) for convergence (impractical for natural image datasets). Attention exhibits 6-9× higher influence malleability than ReLU networks, enabling task alignment but increasing adversarial susceptibility.

Conclusion: Attention’s power and vulnerability share a common origin in its departure from the kernel regime - its data-dependent kernel can align with task structure but also increases susceptibility to adversarial manipulation, explaining both its effectiveness and security concerns.

Abstract: Understanding the theoretical foundations of attention mechanisms remains challenging due to their complex, non-linear dynamics. This work reveals a fundamental trade-off in the learning dynamics of linearized attention. Using a linearized attention mechanism with exact correspondence to a data-dependent Gram-induced kernel, both empirical and theoretical analysis through the Neural Tangent Kernel (NTK) framework shows that linearized attention does not converge to its infinite-width NTK limit, even at large widths. A spectral amplification result establishes this formally: the attention transformation cubes the Gram matrix’s condition number, requiring width $m = Ω(κ^6)$ for convergence, a threshold that exceeds any practical width for natural image datasets. This non-convergence is characterized through influence malleability, the capacity to dynamically alter reliance on training examples. Attention exhibits 6–9$\times$ higher malleability than ReLU networks, with dual implications: its data-dependent kernel can reduce approximation error by aligning with task structure, but this same sensitivity increases susceptibility to adversarial manipulation of training data. These findings suggest that attention’s power and vulnerability share a common origin in its departure from the kernel regime.

Method: Replace engineered priors with learned priors from a foundation model pre-trained on millions of regression tasks. The model performs in-context learning to instantly adapt to each circuit without tuning or retraining. Uses attention mechanism to transfer knowledge across corners by identifying shared circuit physics, combined with automated feature selection (1152D to 48D).

Result: Matches state-of-the-art accuracy with mean MREs as low as 0.11% while requiring zero tuning. Reduces total validation cost by over 10x compared to existing methods.

Conclusion: The foundation model approach breaks the Tuning Barrier in circuit validation by eliminating hyperparameter tuning through in-context learning and knowledge transfer across operating conditions, enabling efficient multi-corner analysis.

Abstract: Yield Multi-Corner Analysis validates circuits across 25+ Process-Voltage-Temperature corners, resulting in a combinatorial simulation cost of $O(K \times N)$ where $K$ denotes corners and $N$ exceeds $10^4$ samples per corner. Existing methods face a fundamental trade-off: simple models achieve automation but fail on nonlinear circuits, while advanced AI models capture complex behaviors but require hours of hyperparameter tuning per design iteration, forming the Tuning Barrier. We break this barrier by replacing engineered priors (i.e., model specifications) with learned priors from a foundation model pre-trained on millions of regression tasks. This model performs in-context learning, instantly adapting to each circuit without tuning or retraining. Its attention mechanism automatically transfers knowledge across corners by identifying shared circuit physics between operating conditions. Combined with an automated feature selector (1152D to 48D), our method matches state-of-the-art accuracy (mean MREs as low as 0.11%) with zero tuning, reducing total validation cost by over $10\times$.

Method: BoSS constructs candidate batches through an ensemble of selection strategies, then selects the batch yielding the highest performance gain. It’s designed to be extensible with new strategies as they emerge.

Result: BoSS outperforms existing oracle strategies, shows state-of-the-art AL strategies still fall short of oracle performance (especially in large-scale datasets with many classes), and suggests ensemble-based approaches might improve AL strategy consistency.

Conclusion: BoSS provides a scalable oracle reference for active learning evaluation, reveals significant gaps in current AL methods, and suggests ensemble approaches could address performance inconsistencies.

Abstract: Active learning (AL) aims to reduce annotation costs while maximizing model performance by iteratively selecting valuable instances. While foundation models have made it easier to identify these instances, existing selection strategies still lack robustness across different models, annotation budgets, and datasets. To highlight the potential weaknesses of existing AL strategies and provide a reference point for research, we explore oracle strategies, i.e., strategies that approximate the optimal selection by accessing ground-truth information unavailable in practical AL scenarios. Current oracle strategies, however, fail to scale effectively to large datasets and complex deep neural networks. To tackle these limitations, we introduce the Best-of-Strategy Selector (BoSS), a scalable oracle strategy designed for large-scale AL scenarios. BoSS constructs a set of candidate batches through an ensemble of selection strategies and then selects the batch yielding the highest performance gain. As an ensemble of selection strategies, BoSS can be easily extended with new state-of-the-art strategies as they emerge, ensuring it remains a reliable oracle strategy in the future. Our evaluation demonstrates that i) BoSS outperforms existing oracle strategies, ii) state-of-the-art AL strategies still fall noticeably short of oracle performance, especially in large-scale datasets with many classes, and iii) one possible solution to counteract the inconsistent performance of AL strategies might be to employ an ensemble-based approach for the selection.

Method: Proposes Zero-Order Sharpness-Aware Minimization (ZO-SAM) that integrates zero-order optimization within SAM framework, requiring only single backpropagation during perturbation and using selective zero-order gradient estimations to reduce computational cost by half.

Result: ZO-SAM reduces backpropagation cost by 50% compared to conventional SAM, lowers gradient variance, stabilizes training, accelerates convergence, and improves model robustness under distribution shift.

Conclusion: ZO-SAM provides an efficient optimization framework for sparse neural network training that addresses computational bottlenecks while maintaining improved convergence and robustness properties.

Abstract: Deep learning models, despite their impressive achievements, suffer from high computational costs and memory requirements, limiting their usability in resource-constrained environments. Sparse neural networks significantly alleviate these constraints by dramatically reducing parameter count and computational overhead. However, existing sparse training methods often experience chaotic and noisy gradient signals, severely hindering convergence and generalization performance, particularly at high sparsity levels. To tackle this critical challenge, we propose Zero-Order Sharpness-Aware Minimization (ZO-SAM), a novel optimization framework that strategically integrates zero-order optimization within the SAM approach. Unlike traditional SAM, ZO-SAM requires only a single backpropagation step during perturbation, selectively utilizing zero-order gradient estimations. This innovative approach reduces the backpropagation computational cost by half compared to conventional SAM, significantly lowering gradient variance and effectively eliminating associated computational overhead. By harnessing SAM’s capacity for identifying flat minima, ZO-SAM stabilizes the training process and accelerates convergence. These efficiency gains are particularly important in sparse training scenarios, where computational cost is the primary bottleneck that limits the practicality of SAM. Moreover, models trained with ZO-SAM exhibit improved robustness under distribution shift, further broadening its practicality in real-world deployments.

Method: Proposes a bootstrapping (BS) framework that links the squeezing effect with the model’s own high-confidence generations (model beliefs). BS-T suppresses high-probability tokens while BS-S removes entire high-confidence generations, jointly suppressing both target responses and model beliefs.

Result: Extensive experiments across diverse benchmarks with various model families confirm the approach’s effectiveness in achieving more thorough forgetting while preserving utility compared to previous methods.

Conclusion: The bootstrapping framework addresses fundamental limitations in LLM unlearning by explicitly countering the squeezing effect through model beliefs, enabling more robust and thorough content removal while maintaining model utility.

Abstract: Large language models trained on vast corpora inherently risk memorizing sensitive or harmful content, which may later resurface in their outputs. Prevailing unlearning methods generally rely on gradient ascent and its variants to lower the probability of specific target responses. However, we find that this strategy induces a critical side effect: probability mass is redistributed into high-likelihood regions, often corresponding to semantically related rephrasings of the targets. We refer to this as the squeezing effect, which explains why many methods yield merely spurious unlearning, a problem further obscured by automated metrics (e.g., ROUGE, truth ratio) that misreport actual success. To address this, we propose a bootstrapping (BS) framework that explicitly links the squeezing effect with the model’s own high-confidence generations, namely its model beliefs. Since model beliefs inherently capture the very high-likelihood regions where probability mass is squeezed, incorporating them into the unlearning objective directly counters the squeezing effect. By jointly suppressing both target responses and model beliefs, BS-T (token) attenuates high-probability tokens, whereas BS-S (sequence) removes entire high-confidence generations, together achieving more thorough forgetting while preserving utility. Extensive experiments across diverse benchmarks with various model families confirm the effectiveness of our approach.

Method: MXNorm leverages the block scales calculated during MXFP8 casting to estimate the RMS value, eliminating the need for separate high-precision reduction operations. This enables a 32x decrease in reduction size for normalization.

Result: Validation on Llama 3 models (125M, 1B, 8B parameters) shows minimal training accuracy loss compared to RMSNorm baseline. Practical kernel speedups of up to 2.4x over RMSNorm using torch.compile, translating to 1.3% speedup in Llama 3 8B transformer layers with MXFP8 and 2.6% with NVFP4.

Conclusion: MXNorm effectively addresses the normalization bottleneck in low-precision training by reusing existing block scale information, achieving significant speedups while maintaining model accuracy.

Abstract: Matrix multiplication performance has long been the major bottleneck to scaling deep learning workloads, which has stimulated the design of new accelerators that use increasingly low-precision number formats. However, improvements in matrix multiplication performance have far outstripped improvements in performance on reductions and elementwise computations, which are still being performed in higher precision. In this work, we propose MXNorm, a drop-in replacement for RMSNorm that estimates the RMS using only the block scales calculated as part of the MXFP8 cast and enables a 32x decrease in the size of reduction needed for normalization. We validate our approximation method on pre-training of Llama 3 models of 125M, 1B and 8B parameters, finding minimal loss of training accuracy compared to a baseline using RMSNorm with MXFP8 matmuls. We also show practical kernel speedups using only torch.compile of up to 2.4x for MXNorm over RMSNorm, corresponding to a 1.3% speedup in Llama 3 8B transformer layers in MXFP8 and a 2.6% speedup in NVFP4.

Method: The approach identifies that privacy vulnerability exists in a small fraction of weights, scores these critical weights, and selectively rewinds only those weights for fine-tuning rather than discarding entire neurons or retraining the whole network.

Result: Extensive experiments show the method exhibits superior resilience against Membership Inference Attacks in most cases while maintaining model utility.

Conclusion: Selective weight rewinding for privacy-critical weights provides an effective approach to membership privacy preservation that balances privacy protection with utility maintenance.

Abstract: Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.

Method: The paper evaluates various general-purpose self-supervised learning methods on their ability to learn physics-grounded representations useful for downstream scientific tasks. They compare methods that learn in latent space (like Joint Embedding Predictive Architectures - JEPAs) against those optimizing pixel-level prediction objectives.

Result: Surprisingly, not all methods designed for physical modeling outperform generic self-supervised learning methods. Methods that learn in latent space (e.g., JEPAs) outperform those optimizing pixel-level prediction objectives on downstream scientific tasks like physical parameter estimation.

Conclusion: Downstream scientific tasks provide better evaluation of physical relevance than next-frame prediction. Latent-space learning methods are more effective for learning physics-grounded representations useful for scientific applications.

Abstract: Machine learning approaches to spatiotemporal physical systems have primarily focused on next-frame prediction, with the goal of learning an accurate emulator for the system’s evolution in time. However, these emulators are computationally expensive to train and are subject to performance pitfalls, such as compounding errors during autoregressive rollout. In this work, we take a different perspective and look at scientific tasks further downstream of predicting the next frame, such as estimation of a system’s governing physical parameters. Accuracy on these tasks offers a uniquely quantifiable glimpse into the physical relevance of the representations of these models. We evaluate the effectiveness of general-purpose self-supervised methods in learning physics-grounded representations that are useful for downstream scientific tasks. Surprisingly, we find that not all methods designed for physical modeling outperform generic self-supervised learning methods on these tasks, and methods that learn in the latent space (e.g., joint embedding predictive architectures, or JEPAs) outperform those optimizing pixel-level prediction objectives. Code is available at https://github.com/helenqu/physical-representation-learning.

Method: PhysMoDPO integrates WBC into the training pipeline and uses Direct Preference Optimization with physics-based and task-specific rewards to assign preferences to synthesized trajectories, optimizing the diffusion model to produce motions that remain compliant with both physics and original text instructions after WBC conversion.

Result: Experiments on text-to-motion and spatial control tasks show consistent improvements in physical realism and task-related metrics on simulated robots. The method also enables significant improvements for zero-shot motion transfer in simulation and real-world deployment on a G1 humanoid robot.

Conclusion: PhysMoDPO successfully bridges the gap between text-conditioned motion generation and physics-compliant robot control by optimizing diffusion models to produce motions that maintain fidelity to original instructions while being physically executable, enabling better real-world robot deployment.

Abstract: Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.

Method: Conducted empirical study across various tasks and model backbones, analyzing performance variation with update magnitude (number of demonstrations), example ordering, and selection policies. Compared many-shot prompting with alternative strategies like Dynamic ICL and Reinforced ICL that control information injection and behavior constraints.

Result: Many-shot prompting is effective for structured tasks where demonstrations provide high information gain, but is highly sensitive to selection strategy and shows limited benefits for open-ended generation tasks. Characterized practical limits of prompt-based test-time adaptation.

Conclusion: The study outlines when input-space updates (prompt-based test-time adaptation) are beneficial versus harmful, providing guidance on the practical limitations of many-shot prompting for LLM adaptation.

Abstract: Test-time adaptation enables large language models (LLMs) to modify their behavior at inference without updating model parameters. A common approach is many-shot prompting, where large numbers of in-context learning (ICL) examples are injected as an input-space test-time update. Although performance can improve as more demonstrations are added, the reliability and limits of this update mechanism remain poorly understood, particularly for open-source models. We present an empirical study of many-shot prompting across tasks and model backbones, analyzing how performance varies with update magnitude, example ordering, and selection policy. We further study Dynamic and Reinforced ICL as alternative test-time update strategies that control which information is injected and how it constrains model behavior. We find that many-shot prompting is effective for structured tasks where demonstrations provide high information gain, but is highly sensitive to selection strategy and often shows limited benefits for open-ended generation tasks. Overall, we characterize the practical limits of prompt-based test-time adaptation and outline when input-space updates are beneficial versus harmful.

Method: Advocates for integrating causal methods into ML systems. Examines how causality can be practically integrated into both traditional ML and foundation models to align competing objectives. Draws on existing applications where causality has successfully aligned goals like fairness and accuracy or privacy and robustness.

Result: Causal approaches offer solutions to enhance reliability and interpretability of ML systems. They provide a framework for balancing multiple competing objectives in trustworthy ML and foundation models, addressing conflicts that arise when objectives are pursued in isolation.

Conclusion: Causal frameworks are essential for developing more accountable and ethically sound AI systems. While challenges and limitations exist in adopting causal methods, they offer significant opportunities for creating trustworthy ML systems that can navigate complex trade-offs among competing objectives.

Abstract: Ensuring trustworthiness in machine learning (ML) systems is crucial as they become increasingly embedded in high-stakes domains. This paper advocates for integrating causal methods into machine learning to navigate the trade-offs among key principles of trustworthy ML, including fairness, privacy, robustness, accuracy, and explainability. While these objectives should ideally be satisfied simultaneously, they are often addressed in isolation, leading to conflicts and suboptimal solutions. Drawing on existing applications of causality in ML that successfully align goals such as fairness and accuracy or privacy and robustness, this paper argues that a causal approach is essential for balancing multiple competing objectives in both trustworthy ML and foundation models. Beyond highlighting these trade-offs, we examine how causality can be practically integrated into ML and foundation models, offering solutions to enhance their reliability and interpretability. Finally, we discuss the challenges, limitations, and opportunities in adopting causal frameworks, paving the way for more accountable and ethically sound AI systems.

Method: Introduces Guided Policy Optimization (GPO) framework that co-trains a guider (with access to privileged information) and a learner (trained via imitation learning). The guider ensures alignment with the learner’s policy while leveraging extra information available in simulations.

Result: Theoretical analysis shows GPO achieves optimality comparable to direct RL. Empirical evaluations demonstrate strong performance across various tasks including continuous control with partial observability and noise, and memory-based challenges, significantly outperforming existing methods.

Conclusion: GPO effectively leverages privileged information in simulations to improve RL in partially observable environments, overcoming limitations of existing approaches through a co-training framework of guider and learner.

Abstract: Reinforcement Learning (RL) in partially observable environments poses significant challenges due to the complexity of learning under uncertainty. While additional information, such as that available in simulations, can enhance training, effectively leveraging it remains an open problem. To address this, we introduce Guided Policy Optimization (GPO), a framework that co-trains a guider and a learner. The guider takes advantage of privileged information while ensuring alignment with the learner’s policy that is primarily trained via imitation learning. We theoretically demonstrate that this learning scheme achieves optimality comparable to direct RL, thereby overcoming key limitations inherent in existing approaches. Empirical evaluations show strong performance of GPO across various tasks, including continuous control with partial observability and noise, and memory-based challenges, significantly outperforming existing methods.

Method: Two key improvements: 1) Leverage uncertainty estimates of base policy to focus exploration on uncertain regions, 2) Modify off-policy residual learning to observe base actions and handle stochastic policies better. Tested with Gaussian-based and Diffusion-based stochastic base policies.

Result: Significantly outperforms state-of-the-art finetuning methods, demo-augmented RL methods, and other residual RL methods on Robosuite and D4RL benchmarks. Demonstrates robust zero-shot sim-to-real transfer in real-world deployment.

Conclusion: The proposed uncertainty-aware residual RL method effectively adapts stochastic pretrained policies with improved sample efficiency and handles sparse rewards better than existing approaches.

Abstract: Residual Reinforcement Learning (RL) is a popular approach for adapting pretrained policies by learning a lightweight residual policy that provides corrective actions. While Residual RL is more sample-efficient than finetuning the entire base policy, existing methods struggle with sparse rewards and are designed for deterministic base policies. We propose two improvements to Residual RL that further enhance its sample efficiency and make it suitable for stochastic base policies. First, we leverage uncertainty estimates of the base policy to focus exploration on regions in which the base policy is not confident. Second, we propose a simple modification to off-policy residual learning that allows it to observe base actions and better handle stochastic base policies. We evaluate our method with both Gaussian-based and Diffusion-based stochastic base policies on tasks from Robosuite and D4RL, and compare against state-of-the-art finetuning methods, demo-augmented RL methods, and other residual RL methods. Our algorithm significantly outperforms existing baselines in a variety of simulation benchmark environments. We also deploy our learned polices in the real world to demonstrate their robustness with zero-shot sim-to-real transfer. Paper homepage : lakshitadodeja.github.io/uncertainty-aware-residual-rl/

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2401.02739: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2401.02739&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper abstract

Result: Cannot determine results without access to paper abstract

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2401.06279: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2401.06279&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to abstract fetch failure

Result: Unable to determine results due to abstract fetch failure

Conclusion: Unable to determine conclusion due to abstract fetch failure

Abstract: Failed to fetch summary for 2402.02017: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2402.02017&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2407.20912: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.20912&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2409.03658: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.03658&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2412.13852: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.13852&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2502.13900: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.13900&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2505.00818: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.00818&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2505.13702: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.13702&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2505.16736: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.16736&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to API rate limiting

Result: No results available - arXiv API returned HTTP 429 error indicating too many requests

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper content

Abstract: Failed to fetch summary for 2507.05914: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.05914&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2508.00304: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.00304&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2508.07370: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.07370&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2509.16447: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.16447&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2509.20276: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.20276&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2509.20786: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.20786&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable.

Result: Cannot determine results as paper content is unavailable.

Conclusion: Cannot draw conclusions about the paper due to unavailability of content.

Abstract: Failed to fetch summary for 2509.21619: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21619&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to draw conclusions due to access error

Abstract: Failed to fetch summary for 2509.23313: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23313&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content.

Result: Cannot determine results due to inability to access paper content.

Conclusion: Cannot draw conclusions about the paper’s content due to access limitations.

Abstract: Failed to fetch summary for 2509.24868: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24868&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2511.13421: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.13421&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable due to rate limiting from arXiv API

Result: Cannot determine results as abstract is unavailable due to rate limiting from arXiv API

Conclusion: Cannot determine conclusion as abstract is unavailable due to rate limiting from arXiv API

Abstract: Failed to fetch summary for 2512.14873: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.14873&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval error

Result: Unable to determine results due to retrieval error

Conclusion: Unable to determine conclusion due to retrieval error

Abstract: Failed to fetch summary for 2512.22587: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.22587&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting.

Result: Cannot determine results as paper content is unavailable due to API rate limiting.

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting.

Abstract: Failed to fetch summary for 2602.04096: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.04096&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.15510: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.15510&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.10301: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10301&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2603.11139: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.11139&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2208.13701: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2208.13701&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2303.07287: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2303.07287&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2407.15693: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.15693&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is not accessible.

Result: Cannot determine results as paper content is not accessible.

Conclusion: Cannot draw conclusions about the paper due to unavailability of content.

Abstract: Failed to fetch summary for 2410.03191: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.03191&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2505.10628: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.10628&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2507.19861: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.19861&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2508.06550: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.06550&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2509.06703: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.06703&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as paper content could not be retrieved

Result: Unable to determine results as paper content could not be retrieved

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2510.01112: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01112&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.01930: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01930&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2511.02620: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.02620&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.00299: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00299&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing abstract content

Result: Cannot determine results due to missing abstract content

Conclusion: Cannot draw conclusions due to missing abstract content

Abstract: Failed to fetch summary for 2601.04478: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04478&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to abstract fetch failure

Result: Unable to determine results due to abstract fetch failure

Conclusion: Unable to determine conclusion due to abstract fetch failure

Abstract: Failed to fetch summary for 2601.17907: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17907&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2602.10541: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10541&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.21130: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21130&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: DIALECTIC uses an LLM-based multi-agent system that: 1) Gathers factual knowledge about startups and organizes it into a hierarchical question tree, 2) Synthesizes facts into natural-language arguments for and against investment, 3) Conducts iterative critique and refinement through simulated debates to surface the most convincing arguments, and 4) Produces numeric decision scores for ranking opportunities.

Result: The system was evaluated through backtesting on real investment opportunities from five VC funds. DIALECTIC matches the precision of human VCs in predicting startup success, demonstrating its effectiveness as a screening tool.

Conclusion: DIALECTIC provides an effective automated solution for venture capital screening that can help investors efficiently prioritize opportunities while maintaining human-level predictive accuracy, potentially easing the bandwidth constraints faced by VCs.

Abstract: Venture capital (VC) investors face a large number of investment opportunities but only invest in few of these, with even fewer ending up successful. Early-stage screening of opportunities is often limited by investor bandwidth, demanding tradeoffs between evaluation diligence and number of opportunities assessed. To ease this tradeoff, we introduce DIALECTIC, an LLM-based multi-agent system for startup evaluation. DIALECTIC first gathers factual knowledge about a startup and organizes these facts into a hierarchical question tree. It then synthesizes the facts into natural-language arguments for and against an investment and iteratively critiques and refines these arguments through a simulated debate, which surfaces only the most convincing arguments. Our system also produces numeric decision scores that allow investors to rank and thus efficiently prioritize opportunities. We evaluate DIALECTIC through backtesting on real investment opportunities aggregated from five VC funds, showing that DIALECTIC matches the precision of human VCs in predicting startup success.

Method: Proposes CoMAM (Collaborative Reinforcement Learning Framework for Multi-Agent Memory Systems) that models agent execution as a sequential Markov decision process to embed inter-agent dependencies. Uses group-level ranking consistency between local and global rewards to quantify each agent’s contribution, then integrates these as adaptive weights to assign global credit and combine local-global rewards for joint optimization.

Result: Experiments show CoMAM outperforms leading memory systems, validating the efficacy of the proposed collaborative reinforcement learning approach for joint optimization of multi-agent memory systems.

Conclusion: The collaborative reinforcement learning framework successfully addresses the limitation of independent agent optimization in memory systems, demonstrating that joint optimization through embedded inter-agent dependencies and integrated reward mechanisms leads to superior global system performance.

Abstract: Memory systems are crucial to personalized LLMs by mitigating the context window limitation in capturing long-term user-LLM conversations. Typically, such systems leverage multiple agents to handle multi-granular memory construction and personalized memory retrieval tasks. To optimize the system, existing methods focus on specializing agents on their local tasks independently via prompt engineering or fine-tuning. However, they overlook cross-agent collaboration, where independent optimization on local agents hardly guarantees the global system performance. To address this issue, we propose a Collaborative Reinforcement Learning Framework for Multi-Agent Memory Systems (CoMAM), jointly optimizing local agents to facilitate collaboration. Specifically, we regularize agents’ execution as a sequential Markov decision process (MDP) to embed inter-agent dependencies into the state transition, yielding both local task rewards (e.g., information coverage for memory construction) and global rewards (i.e., query-answer accuracy). Then, we quantify each agent’s contribution via group-level ranking consistency between local and global rewards, treating them as adaptive weights to assign global credit and integrate local-global rewards. Each agent is optimized by these integrated rewards, aligning local improvements with the global performance. Experiments show CoMAM outperforms leading memory systems, validating the efficacy of our proposed collaborative reinforcement learning for joint optimization.

Method: Constitutional Multi-Agent Governance (CMAG) - a two-stage framework that sits between LLM policy compiler and networked agent population, combining hard constraint filtering with soft penalized-utility optimization. Introduces Ethical Cooperation Score (ECS) that multiplies cooperation, autonomy, integrity, and fairness metrics.

Result: CMAG achieves ECS of 0.741 (14.9% improvement over unconstrained optimization), preserves autonomy at 0.985 and integrity at 0.995, with cooperation reduction to 0.770. Unconstrained optimization achieves highest raw cooperation (0.873) but lowest ECS (0.645) due to autonomy erosion and fairness degradation.

Conclusion: Cooperation is not inherently desirable without governance; constitutional constraints are necessary to ensure LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria. CMAG dominates the cooperation-autonomy trade-off space.

Abstract: Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.

Method: Computational theory of status based on appropriateness theory, validated through simulations of LLM-based agents in the Concordia framework with experimental manipulation of social visibility within naturalistic agent daily routines.

Result: Social interactions transform functional demand into status-seeking behavior, producing price run-ups and positive price elasticity (Veblen effects) for both real luxury items and synthetic goods, with influencer agents driving endogenous formation of distinct subcultures through targeted sanctioning.

Conclusion: Provides a generative bridge between micro-level cognition and macro-level economic/sociological phenomena, offering a new methodology for forecasting cultural convention emergence from interaction.

Abstract: Status signaling drives human behavior and the allocation of scarce resources such as mating opportunities, yet the generative mechanisms governing how specific goods, signals, or behaviors acquire prestige remain a puzzle. Classical frameworks, such as Costly Signaling Theory, treat preferences as fixed and struggle to explain how semiotic meaning changes based on context or drifts dynamically over time, occasionally reaching tipping points. In this work, we propose a computational theory of status grounded in the theory of appropriateness, positing that status symbols emerge endogenously through a feedback loop of social observation and predictive pattern completion. We validate this theory using simulations of groups of Large Language Model (LLM)-based agents in the Concordia framework. By experimentally manipulating social visibility within naturalistic agent daily routines, we demonstrate that social interactions transform functional demand into status-seeking behavior. We observe the emergence of price run-ups and positive price elasticity (Veblen effects) for both real-world luxury items and procedurally generated synthetic goods, ruling out pretraining bias as the sole driver. Furthermore, we demonstrate that “influencer” agents can drive the endogenous formation of distinct subcultures through targeted sanctioning, and find that similar social influence effects generalize to non-monetary signaling behaviors. This work provides a generative bridge between micro-level cognition and macro-level economic and sociological phenomena, offering a new methodology for forecasting how cultural conventions emerge from interaction.

Method: Proposes OmniForcing framework that distills offline dual-stream bidirectional diffusion models into streaming autoregressive generators. Key innovations include: 1) Asymmetric Block-Causal Alignment with zero-truncation Global Prefix to handle modality asymmetry, 2) Audio Sink Token mechanism with Identity RoPE constraint to address audio token sparsity, 3) Joint Self-Forcing Distillation to correct cumulative cross-modal errors from exposure bias, and 4) modality-independent rolling KV-cache inference scheme.

Result: Achieves state-of-the-art streaming generation at ~25 FPS on a single GPU while maintaining multi-modal synchronization and visual quality comparable to the bidirectional teacher model. Successfully addresses training instability issues from modality asymmetry and token sparsity.

Conclusion: OmniForcing enables real-time, high-fidelity audio-visual generation by effectively distilling bidirectional diffusion models into efficient streaming autoregressive generators, solving key challenges in multi-modal temporal alignment and training stability.

Abstract: Recent joint audio-visual diffusion models achieve remarkable generation quality but suffer from high latency due to their bidirectional attention dependencies, hindering real-time applications. We propose OmniForcing, the first framework to distill an offline, dual-stream bidirectional diffusion model into a high-fidelity streaming autoregressive generator. However, naively applying causal distillation to such dual-stream architectures triggers severe training instability, due to the extreme temporal asymmetry between modalities and the resulting token sparsity. We address the inherent information density gap by introducing an Asymmetric Block-Causal Alignment with a zero-truncation Global Prefix that prevents multi-modal synchronization drift. The gradient explosion caused by extreme audio token sparsity during the causal shift is further resolved through an Audio Sink Token mechanism equipped with an Identity RoPE constraint. Finally, a Joint Self-Forcing Distillation paradigm enables the model to dynamically self-correct cumulative cross-modal errors from exposure bias during long rollouts. Empowered by a modality-independent rolling KV-cache inference scheme, OmniForcing achieves state-of-the-art streaming generation at $\sim$25 FPS on a single GPU, maintaining multi-modal synchronization and visual quality on par with the bidirectional teacher.\textbf{Project Page:} \href{https://omniforcing.com}{https://omniforcing.com}

Method: Proposes MamTra, an interleaved Mamba-Transformer framework that leverages Mamba’s efficiency and Transformer’s modeling capability. Uses novel knowledge transfer strategies to distill insights from pretrained Transformers into the hybrid architecture, avoiding expensive training from scratch.

Result: MamTra reduces inference VRAM usage by up to 34% without compromising speech fidelity, even when trained on only 2% of the original training dataset. Systematic experiments identified optimal hybrid configurations.

Conclusion: The hybrid Mamba-Transformer approach successfully balances computational efficiency with modeling capability for text-to-speech synthesis, making LLM-based TTS more practical for real-world applications.

Abstract: Despite the remarkable quality of LLM-based text-to-speech systems, their reliance on autoregressive Transformers leads to quadratic computational complexity, which severely limits practical applications. Linear-time alternatives, notably Mamba, offer a potential remedy; however, they often sacrifice the global context essential for expressive synthesis. In this paper, we propose MamTra, an interleaved Mamba-Transformer framework designed to leverage the advantages of Mamba’s efficiency and Transformers’ modeling capability. We also introduce novel knowledge transfer strategies to distill insights from a pretrained Transformer into our hybrid architecture, thereby bypassing the prohibitive costs of training from scratch. Systematic experiments identify the optimal hybrid configuration, and demonstrate that MamTra reduces inference VRAM usage by up to 34% without compromising speech fidelity - even trained on only 2% of the original training dataset. Audio samples are available at https://mamtratts.github.io.

Method: Proposes a diffusion-based RIR completion method using signal-prediction conditioned on ISM-simulated direct-path and early reflections, incorporating classifier-free guidance to steer generation toward target distribution from physically realistic RIRs simulated with Treble SDK.

Result: Objective evaluation shows the method outperforms state-of-the-art baseline in early RIR completion and energy decay curve reconstruction.

Conclusion: The diffusion-based approach successfully generates realistic RIRs from simulated early reflections without duration constraints, bridging the gap between efficient geometric simulation and measured RIR realism.

Abstract: Room impulse responses (RIRs) are fundamental to audio data augmentation, acoustic signal processing, and immersive audio rendering. While geometric simulators such as the image source method (ISM) can efficiently generate early reflections, they lack the realism of measured RIRs due to missing acoustic wave effects. We propose a diffusion-based RIR completion method using signal-prediction conditioned on ISM-simulated direct-path and early reflections. Unlike state-of-the-art methods, our approach imposes no fixed duration constraint on the input early reflections. We further incorporate classifier-free guidance to steer generation toward a target distribution learned from physically realistic RIRs simulated with the Treble SDK. Objective evaluation demonstrates that the proposed method outperforms a state-of-the-art baseline in early RIR completion and energy decay curve reconstruction.

Method: Analyze S3M representations to show compositional encoding of phonological information from neighboring phones, demonstrating orthogonality between relative positions and emergence of implicit phonetic boundaries.

Result: Found that S3Ms compositionally encode phonological vectors from previous, current, and next phones within single frame representations, with orthogonal position encoding and implicit boundary detection.

Conclusion: S3M representations exhibit sophisticated compositional structure that encodes contextual phonological information, advancing understanding of context-dependent speech representations.

Abstract: Transformer-based self-supervised speech models (S3Ms) are often described as contextualized, yet what this entails remains unclear. Here, we focus on how a single frame-level S3M representation can encode phones and their surrounding context. Prior work has shown that S3Ms represent phones compositionally; for example, phonological vectors such as voicing, bilabiality, and nasality vectors are superposed in the S3M representation of [m]. We extend this view by proposing that phonological information from a sequence of neighboring phones is also compositionally encoded in a single frame, such that vectors corresponding to previous, current, and next phones are superposed within a single frame-level representation. We show that this structure has several properties, including orthogonality between relative positions, and emergence of implicit phonetic boundaries. Together, our findings advance our understanding of context-dependent S3M representations.

Method: The authors derive statistical bounds on PCC and MSE based on subjective vote variance. They introduce a binomial-based model for subjective votes (BinoVotes) that leads to a mean opinion score model (BinoMOS) with desirable properties. The model accounts for discrete MOS values and their dependence on vote counts.

Result: The BinoMOS model provides vote variance information needed for PCC and MSE bounds. Validation against 18 subjective tests shows the modeling yields bounds that agree well with those derived directly from data. The approach allows setting realistic expectations for objective-subjective comparisons.

Conclusion: The paper provides a framework for setting realistic bounds on objective quality metrics by accounting for subjective test noise, enabling better evaluation of multimedia quality estimators even when vote variance information is unavailable.

Abstract: Objective estimators of multimedia quality are often judged by comparing estimates with subjective “truth data,” most often via Pearson correlation coefficient (PCC) or mean-squared error (MSE). But subjective test results contain noise, so striving for a PCC of 1.0 or an MSE of 0.0 is neither realistic nor repeatable. Numerous efforts have been made to acknowledge and appropriately accommodate subjective test noise in objective-subjective comparisons, typically resulting in new analysis frameworks and figures-of-merit. We take a different approach. By making only basic assumptions, we derive bounds on PCC and MSE that can be expected for a subjective test. Consistent with intuition, these bounds are functions of subjective vote variance. When a subjective test includes vote variance information, the calculation of the bounds is easy, and in this case we say the resulting bounds are “fully data-driven.” We provide two options for calculating bounds in cases where vote variance information is not available. One option is to use vote variance information from other subjective tests that do provide such information, and the second option is to use a model for subjective votes. Thus we introduce a binomial-based model for subjective votes (BinoVotes) that naturally leads to a mean opinion score (MOS) model, named BinoMOS, with multiple unique desirable properties. BinoMOS reproduces the discrete nature of MOS values and its dependence on the number of votes per file. This modeling provides vote variance information required by the PCC and MSE bounds and we compare this modeling with data from 18 subjective tests. The modeling yields PCC and MSE bounds that agree very well with those found from the data directly. These results allow one to set expectations for the PCC and MSE that might be achieved for any subjective test, even those where vote variance information is not available.

Method: Proposes GTCRN lightweight speech enhancement model to guide TSE. Two extensions: LGTSE (noise-agnostic enrollment guidance by denoising input before context interaction) and D-LGTSE (uses denoised speech as additional noisy input during training). Two-stage training: GTCRN enhancement-guided pre-training followed by joint fine-tuning.

Result: Experiments on Libri2Mix dataset show significant improvements: 0.89 dB in SISDR, 0.16 in PESQ, and 1.97% in STOI compared to baseline methods.

Conclusion: The proposed approach effectively improves target speech extraction performance in noisy multi-speaker scenarios through guided enhancement and robust training strategies.

Abstract: Target speech extraction (TSE) has achieved strong performance in relatively simple conditions such as one-speaker-plus-noise and two-speaker mixtures, but its performance remains unsatisfactory in noisy multi-speaker scenarios. To address this issue, we introduce a lightweight speech enhancement model, GTCRN, to better guide TSE in noisy environments. Building on our competitive previous speaker embedding/encoder-free framework SEF-PNet, we propose two extensions: LGTSE and D-LGTSE. LGTSE incorporates noise-agnostic enrollment guidance by denoising the input noisy speech before context interaction with enrollment speech, thereby reducing noise interference. D-LGTSE further improves system robustness against speech distortion by leveraging denoised speech as an additional noisy input during training, expanding the dynamic range of noisy conditions and enabling the model to directly learn from distorted signals. Furthermore, we propose a two-stage training strategy, first with GTCRN enhancement-guided pre-training and then joint fine-tuning, to fully exploit model potential.Experiments on the Libri2Mix dataset demonstrate significant improvements of 0.89 dB in SISDR, 0.16 in PESQ, and 1.97% in STOI, validating the effectiveness of our approach.

Method: MAGE uses a scarcity-aware coarse-to-fine masking strategy that prioritizes frequent tokens in early steps and rare tokens in later refinements. It includes a lightweight corrector module to detect low-confidence predictions and re-mask them for refinement. Built on BigCodec and finetuned from Qwen2.5-0.5B, it’s reduced to 200M parameters through selective layer retention.

Result: MAGE achieves state-of-the-art perceptual quality on DNS Challenge and noisy LibriSpeech datasets, significantly reduces word error rate for downstream recognition, and outperforms larger baselines despite having only 200M parameters.

Conclusion: MAGE demonstrates that compact masked generative models with intelligent masking strategies and correction mechanisms can achieve superior speech enhancement quality while maintaining efficiency.

Abstract: Speech enhancement remains challenging due to the trade-off between efficiency and perceptual quality. In this paper, we introduce MAGE, a Masked Audio Generative Enhancer that advances generative speech enhancement through a compact and robust design. Unlike prior masked generative models with random masking, MAGE employs a scarcity-aware coarse-to-fine masking strategy that prioritizes frequent tokens in early steps and rare tokens in later refinements, improving efficiency and generalization. We also propose a lightweight corrector module that further stabilizes inference by detecting low-confidence predictions and re-masking them for refinement. Built on BigCodec and finetuned from Qwen2.5-0.5B, MAGE is reduced to 200M parameters through selective layer retention. Experiments on DNS Challenge and noisy LibriSpeech show that MAGE achieves state-of-the-art perceptual quality and significantly reduces word error rate for downstream recognition, outperforming larger baselines. Audio examples are available at https://hieugiaosu.github.io/MAGE/.

Method: Proposes new framework for data creation and research methodology that incorporates communication channel effects, focusing on creating more realistic datasets that simulate real-world transmission scenarios rather than just using raw deepfake audio.

Result: Improved deepfake detection accuracy by 39% in robust lab setups and 57% on real-world benchmarks. Demonstrated that better datasets have bigger impact on accuracy than using larger state-of-the-art models over smaller ones.

Conclusion: Scientific community should prioritize comprehensive data collection programs over training larger models with higher computational demands, as realistic datasets accounting for communication channels are more crucial for effective real-world deepfake detection.

Abstract: While the technologies empowering malicious audio deepfakes have dramatically evolved in recent years due to generative AI advances, the same cannot be said of global research into spoofing (deepfake) countermeasures. This paper highlights how current deepfake datasets and research methodologies led to systems that failed to generalize to real world application. The main reason is due to the difference between raw deepfake audio, and deepfake audio that has been presented through a communication channel, e.g. by phone. We propose a new framework for data creation and research methodology, allowing for the development of spoofing countermeasures that would be more effective in real-world scenarios. By following the guidelines outlined here we improved deepfake detection accuracy by 39% in more robust and realistic lab setups, and by 57% on a real-world benchmark. We also demonstrate how improvement in datasets would have a bigger impact on deepfake detection accuracy than the choice of larger SOTA models would over smaller models; that is, it would be more important for the scientific community to make greater investment on comprehensive data collection programs than to simply train larger models with higher computational demands.

Method: Proposes DSN with static and dynamic components, targeting common neural network layers (grouped RNN units, multi-head attention, convolutional, fully connected). Uses a policy module for frame-wise dynamic part control and Metric-Guided Training to explicitly guide the policy module in assessing speech quality.

Result: DSN achieves comparable enhancement performance to state-of-the-art lightweight baseline while using only 73% of its computational load on average. Dynamic component usage ratios show appropriate resource allocation based on input signal distortion severity.

Conclusion: The proposed DSN with MGT effectively reduces computational complexity while maintaining speech enhancement performance, demonstrating intelligent resource allocation based on input quality.

Abstract: To further reduce the complexity of lightweight speech enhancement models, we introduce a gating-based Dynamically Slimmable Network (DSN). The DSN comprises static and dynamic components. For architecture-independent applicability, we introduce distinct dynamic structures targeting the commonly used components, namely, grouped recurrent neural network units, multi-head attention, convolutional, and fully connected layers. A policy module adaptively governs the use of dynamic parts at a frame-wise resolution according to the input signal quality, controlling computational load. We further propose Metric-Guided Training (MGT) to explicitly guide the policy module in assessing input speech quality. Experimental results demonstrate that the DSN achieves comparable enhancement performance in instrumental metrics to the state-of-the-art lightweight baseline, while using only 73% of its computational load on average. Evaluations of dynamic component usage ratios indicate that the MGT-DSN can appropriately allocate network resources according to the severity of input signal distortion.

Method: Extends LGTSE with TripleC (Cross-Condition Consistency) learning strategy that enforces consistent speech extraction across different acoustic conditions. Uses parallel universal training scheme organizing batches containing multiple scenarios for the same target speaker, allowing easier cases to assist harder ones.

Result: Experimental results on Libri2Mix three-condition tasks show LGTSE with TripleC learning achieves superior performance over condition-specific models, demonstrating strong generalization across diverse scenarios.

Conclusion: The proposed approach enables robust universal deployment in real-world speech applications by effectively leveraging diverse training data and enforcing cross-condition consistency, outperforming specialized models for each condition.

Abstract: In our recent work, we proposed Lightweight Speech Enhancement Guided Target Speech Extraction (LGTSE) and demonstrated its effectiveness in multi-speaker-plus-noise scenarios. However, real-world applications often involve more diverse and complex conditions, such as one-speaker-plus-noise or two-speaker-without-noise. To address this challenge, we extend LGTSE with a Cross-Condition Consistency learning strategy, termed TripleC Learning. This strategy is first validated under multi-speaker-plus-noise condition and then evaluated for its generalization across diverse scenarios. Moreover, building upon the lightweight front-end denoiser in LGTSE, which can flexibly process both noisy and clean mixtures and shows strong generalization to unseen conditions, we integrate TripleC learning with a proposed parallel universal training scheme that organizes batches containing multiple scenarios for the same target speaker. By enforcing consistent extraction across different conditions, easier cases can assist harder ones, thereby fully exploiting diverse training data and fostering a robust universal model. Experimental results on the Libri2Mix three-condition tasks demonstrate that the proposed LGTSE with TripleC learning achieves superior performance over condition-specific models, highlighting its strong potential for universal deployment in real-world speech applications.

Method: Theoretical modeling of diffusion editing as stochastic transformation that contracts off-manifold perturbations, deriving bounds on watermark signal-to-noise ratio and mutual information along diffusion trajectories, plus empirical evaluation of representative watermarking systems under various diffusion-based editing scenarios.

Result: Diffusion editing can significantly reduce watermark recoverability, even with routine semantic edits, making reliable recovery information-theoretically impossible under certain conditions.

Conclusion: Current watermarking approaches are vulnerable to diffusion-based editing, requiring new design principles for watermarks that remain robust under generative image editing for effective content provenance.

Abstract: Robust invisible watermarks are widely used to support copyright protection, content provenance, and accountability by embedding hidden signals designed to survive common post-processing operations. However, diffusion-based image editing introduces a fundamentally different class of transformations: it injects noise and reconstructs images through a powerful generative prior, often altering semantic content while preserving photorealism. In this paper, we provide a unified theoretical and empirical analysis showing that non-adversarial diffusion editing can unintentionally degrade or remove robust watermarks. We model diffusion editing as a stochastic transformation that progressively contracts off-manifold perturbations, causing the low-amplitude signals used by many watermarking schemes to decay. Our analysis derives bounds on watermark signal-to-noise ratio and mutual information along diffusion trajectories, yielding conditions under which reliable recovery becomes information-theoretically impossible. We further evaluate representative watermarking systems under a range of diffusion-based editing scenarios and strengths. The results indicate that even routine semantic edits can significantly reduce watermark recoverability. Finally, we discuss the implications for content provenance and outline principles for designing watermarking approaches that remain robust under generative image editing.

Method: Analyzed 13 cancer benchmark datasets using 4 CNN architectures across different cancer types. Compared model accuracy on original datasets versus datasets made of cropped background segments containing no clinical information.

Result: CNN models achieved high accuracy (up to 93%) on cropped background segments lacking biomedical information, indicating models learn dataset biases rather than clinically relevant features.

Conclusion: Common ML evaluation practices may lead to unreliable results in cancer pathology, as biases in benchmark datasets are difficult to identify and can mislead researchers about CNN efficacy.

Abstract: Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.

Method: Proposes MSG-LDM, a latent diffusion-based framework with style-structure disentanglement in latent space. Separates modality-specific style features from shared structural representations, models low-frequency anatomical layouts and high-frequency boundary details in multi-scale feature space. Uses style consistency loss and structure-aware loss to improve stability.

Result: Extensive experiments on BraTS2020 and WMH datasets show the method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The code is publicly available.

Conclusion: MSG-LDM effectively addresses anatomical inconsistency and texture degradation in multi-modal MRI translation through latent space disentanglement and multi-scale feature modeling, achieving superior performance in structure reconstruction.

Abstract: Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style–structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at https://github.com/ziyi-start/MSG-LDM.

Method: Multiview deep learning framework using parallel convolutional branches to extract features from multidirectional scleral vessel images (5 gaze directions per participant), refined with Manta Ray Foraging Optimization, and fused via transformer-based cross-view attention for three-class classification and continuous glucose estimation.

Result: Achieved 93.8% overall accuracy for metabolic status classification with AUCs of 0.971, 0.956, and 0.982 for normal, controlled diabetes, and high-glucose diabetes respectively. For FPG estimation: MAE = 6.42 mg/dL, RMSE = 7.91 mg/dL, strong correlation (r = 0.983, R² = 0.966).

Conclusion: Multidirectional scleral vessel imaging with multiview learning is a promising non-invasive approach for glycemic assessment, though multicenter validation is needed before clinical deployment.

Abstract: Regular monitoring of glycemic status is essential for diabetes management, yet conventional blood-based testing can be burdensome for frequent assessment. The sclera contains superficial microvasculature that may exhibit diabetes related alterations and is readily visible on the ocular surface. We propose ScleraGluNet, a multiview deep-learning framework for three-class metabolic status classification (normal, controlled diabetes, and high-glucose diabetes) and continuous fasting plasma glucose (FPG) estimation from multidirectional scleral vessel images. The dataset comprised 445 participants (150/140/155) and 2,225 anterior-segment images acquired from five gaze directions per participant. After vascular enhancement, features were extracted using parallel convolutional branches, refined with Manta Ray Foraging Optimization (MRFO), and fused via transformer-based cross-view attention. Performance was evaluated using subject-wise five-fold cross-validation, with all images from each participant assigned to the same fold. ScleraGluNet achieved 93.8% overall accuracy, with one-vs-rest AUCs of 0.971,0.956, and 0.982 for normal, controlled diabetes, and high-glucose diabetes, respectively. For FPG estimation, the model achieved MAE = 6.42 mg/dL and RMSE = 7.91 mg/dL, with strong correlation to laboratory measurements (r = 0.983; R2 = 0.966). Bland Altman analysis showed a mean bias of +1.45 mg/dL with 95% limits of agreement from -8.33 to +11.23$ mg/dL. These results support multidirectional scleral vessel imaging with multiview learning as a promising noninvasive approach for glycemic assessment, warranting multicenter validation before clinical deployment.

Method: Proposes GLEAM dataset with three modalities: scanning laser ophthalmoscopy fundus images, circumpapillary OCT images, and visual field pattern deviation maps. Develops HAMM framework with hierarchical attentive encoders and light decoders for cross-modal representation learning using masked modeling approach.

Result: First publicly available tri-modal glaucoma dataset with four disease stage annotations. HAMM framework enables effective multimodal integration for glaucoma classification across disease stages.

Conclusion: GLEAM dataset and HAMM framework advance multimodal glaucoma diagnosis by providing comprehensive data and effective cross-modal fusion method for accurate disease staging.

Abstract: We propose glaucoma lesion evaluation and analysis with multimodal imaging (GLEAM), the first publicly available tri-modal glaucoma dataset comprising scanning laser ophthalmoscopy fundus images, circumpapillary OCT images, and visual field pattern deviation maps, annotated with four disease stages, enabling effective exploitation of multimodal complementary information and facilitating accurate diagnosis and treatment across disease stages. To effectively integrate cross-modal information, we propose hierarchical attentive masked modeling (HAMM) for multimodal glaucoma classification. Our framework employs hierarchical attentive encoders and light decoders to focus cross-modal representation learning on the encoder.

Method: Integrate deep learning modules SynthStrip (for skull stripping) and SynthSeg (for tissue segmentation) into the SIENA pipeline. Evaluate three pipeline variants on ADNI and PPMI longitudinal cohorts using three criteria: correlation with clinical/structural decline, scan-order consistency, and runtime.

Result: Replacing skull-stripping module yields most consistent gains: strengthens associations between PBVC and disease progression in ADNI, improves robustness under scan reversal across datasets. Fully integrated pipeline achieves strongest scan-order consistency (reducing error by up to 99.1%). GPU-enabled variants reduce execution time by up to 46% while maintaining comparable CPU runtimes.

Conclusion: Deep learning can meaningfully strengthen established longitudinal atrophy pipelines when used to reinforce their weakest image processing steps. Modular modernization of clinically trusted neuroimaging tools is valuable without sacrificing interpretability.

Abstract: Percentage Brain Volume Change (PBVC) derived from Magnetic Resonance Imaging (MRI) is a widely used biomarker of brain atrophy, with SIENA among the most established methods for its estimation. However, SIENA relies on classical image processing steps, particularly skull stripping and tissue segmentation, whose failures can propagate through the pipeline and bias atrophy estimates. In this work, we examine whether targeted deep learning substitutions can improve SIENA while preserving its established and interpretable framework. To this end, we integrate SynthStrip and SynthSeg into SIENA and evaluate three pipeline variants on the ADNI and PPMI longitudinal cohorts. Performance is assessed using three complementary criteria: correlation with longitudinal clinical and structural decline, scan-order consistency, and end-to-end runtime. Replacing the skull-stripping module yields the most consistent gains: in ADNI, it substantially strengthens associations between PBVC and multiple measures of disease progression relative to the standard SIENA pipeline, while across both datasets it markedly improves robustness under scan reversal. The fully integrated pipeline achieves the strongest scan-order consistency, reducing the error by up to 99.1%. In addition, GPU-enabled variants reduce execution time by up to 46% while maintaining CPU runtimes comparable to standard SIENA. Overall, these findings show that deep learning can meaningfully strengthen established longitudinal atrophy pipelines when used to reinforce their weakest image processing steps. More broadly, this study highlights the value of modularly modernizing clinically trusted neuroimaging tools without sacrificing their interpretability. Code is publicly available at https://github.com/Raciti/Enhanced-SIENA.git.

Method: Two-stage approach: 1) Pre-train diffusion probabilistic model on large diverse brain MRI dataset (~4000 subjects), 2) Fine-tune on small target dataset (20 subjects) with carefully selected learning rates and durations, evaluated on fastMRI experiments and clinical stroke data with blinded reader study.

Result: Models pre-trained on non-FLAIR contrasts and fine-tuned on only 20 FLAIR subjects achieve comparable performance to models trained with much more target data. Moderate fine-tuning with reduced learning rate works best. Blinded reader study shows 2× accelerated reconstructions are non-inferior to standard-of-care for image quality and structural delineation.

Conclusion: Large-scale pre-training with targeted fine-tuning enables diffusion-based MRI reconstruction in data-constrained clinical applications, reducing need for large application-specific datasets while maintaining clinical quality.

Abstract: Purpose: To develop a data-efficient strategy for accelerated MRI reconstruction with Diffusion Probabilistic Generative Models (DPMs) that enables faster scan times in clinical stroke MRI when only limited fully-sampled data samples are available. Methods: Our simple training strategy, inspired by the foundation model paradigm, first trains a DPM on a large, diverse collection of publicly available brain MRI data in fastMRI and then fine-tunes on a small dataset from the target application using carefully selected learning rates and fine-tuning durations. The approach is evaluated on controlled fastMRI experiments and on clinical stroke MRI data with a blinded clinical reader study. Results: DPMs pre-trained on approximately 4000 subjects with non-FLAIR contrasts and fine-tuned on FLAIR data from only 20 target subjects achieve reconstruction performance comparable to models trained with substantially more target-domain FLAIR data across multiple acceleration factors. Experiments reveal that moderate fine-tuning with a reduced learning rate yields improved performance, while insufficient or excessive fine-tuning degrades reconstruction quality. When applied to clinical stroke MRI, a blinded reader study involving two neuroradiologists indicates that images reconstructed using the proposed approach from $2 \times$ accelerated data are non-inferior to standard-of-care in terms of image quality and structural delineation. Conclusion: Large-scale pre-training combined with targeted fine-tuning enables DPM-based MRI reconstruction in data-constrained, accelerated clinical stroke MRI. The proposed approach substantially reduces the need for large application-specific datasets while maintaining clinically acceptable image quality, supporting the use of foundation-inspired diffusion models for accelerated MRI in targeted applications.

Method: Replaces U-Net with Diffusion Transformer operating entirely at 32x downscaled resolution. Adapts pretrained text-to-image multi-step DiT into single-step reconstruction model using three alignment mechanisms: 1) variance-guided reconstruction flow adapting denoising strength to latent uncertainty, 2) self-distillation alignment enforcing consistency with encoder-defined latent geometry for one-step diffusion, and 3) latent-conditioned guidance replacing text prompts with semantically aligned latent conditions for text-free inference.

Result: Achieves state-of-the-art perceptual quality while offering up to 30x faster decoding and drastically lower memory usage than existing diffusion-based codecs. Can reconstruct 2048x2048 images on a 16 GB laptop GPU.

Conclusion: DiT-IC demonstrates that diffusion can operate effectively in compact latent spaces (32x downscaled) without compromising reconstruction quality, making diffusion-based image compression practical through efficient transformer-based architecture and alignment mechanisms.

Abstract: Diffusion-based image compression has recently shown outstanding perceptual fidelity, yet its practicality is hindered by prohibitive sampling overhead and high memory usage. Most existing diffusion codecs employ U-Net architectures, where hierarchical downsampling forces diffusion to operate in shallow latent spaces (typically with only 8x spatial downscaling), resulting in excessive computation. In contrast, conventional VAE-based codecs work in much deeper latent domains (16x - 64x downscaled), motivating a key question: Can diffusion operate effectively in such compact latent spaces without compromising reconstruction quality? To address this, we introduce DiT-IC, an Aligned Diffusion Transformer for Image Compression, which replaces the U-Net with a Diffusion Transformer capable of performing diffusion in latent space entirely at 32x downscaled resolution. DiT-IC adapts a pretrained text-to-image multi-step DiT into a single-step reconstruction model through three key alignment mechanisms: (1) a variance-guided reconstruction flow that adapts denoising strength to latent uncertainty for efficient reconstruction; (2) a self-distillation alignment that enforces consistency with encoder-defined latent geometry to enable one-step diffusion; and (3) a latent-conditioned guidance that replaces text prompts with semantically aligned latent conditions, enabling text-free inference. With these designs, DiT-IC achieves state-of-the-art perceptual quality while offering up to 30x faster decoding and drastically lower memory usage than existing diffusion-based codecs. Remarkably, it can reconstruct 2048x2048 images on a 16 GB laptop GPU.

Method: Uses VAE-compressed image latents fed to a ViT encoder trained with two complementary objectives: 1) reconstruction via ViT-VAE decoder for generative structure, and 2) contrastive learning and image-captioning for semantic features. Joint optimization in shared latent space enables synergy between generative and understanding capabilities.

Result: For generation: significantly outperforms standard CLIP-based encoder (gFID: 1.87 vs. 2.54 on ImageNet). For multimodal understanding: performs comparably with standard CLIP vision encoder (63.3 vs. 61.2 on SeedBench, 59.2 vs. 58.1 on GQA). Shows mutual benefits between generation and understanding.

Conclusion: Unified visual representations for both understanding and generation are feasible and beneficial, with the VAE latent space playing a critical role. The approach demonstrates that generation and understanding tasks can be mutually reinforcing in a shared architecture.

Abstract: This paper presents a family of advanced vision encoder, named OpenVision 3, that learns a single, unified visual representation that can serve both image understanding and image generation. Our core architecture is simple: we feed VAE-compressed image latents to a ViT encoder and train its output to support two complementary roles. First, the encoder output is passed to the ViT-VAE decoder to reconstruct the original image, encouraging the representation to capture generative structure. Second, the same representation is optimized with contrastive learning and image-captioning objectives, strengthening semantic features. By jointly optimizing reconstruction- and semantics-driven signals in a shared latent space, the encoder learns representations that synergize and generalize well across both regimes. We validate this unified design through extensive downstream evaluations with the encoder frozen. For generation, we test it under the RAE framework: ours substantially surpasses the standard CLIP-based encoder (e.g., gFID: 1.87 vs. 2.54 on ImageNet). For multimodal understanding, we plug the encoder into the LLaVA-1.5 and LLaVA-NeXT framework: it performs comparably with a standard CLIP vision encoder (e.g., 63.3 vs. 61.2 on SeedBench, and 59.2 vs. 58.1 on GQA). We provide empirical evidence that generation and understanding are mutually beneficial in our architecture, while further underscoring the critical role of the VAE latent space. We hope this work can spur future research on unified modeling.